PrepWise AI: A Smart Preparation Assistant for Study and Career

PrepWise AI: A Smart Preparation Assistant for Study and Career

Jagadeesh Pujari

Information Science and Engineering SDM College of Engineering and Technology Dharwad, India

Varsha S Jadhav

Information Science and Engineering SDM College of Engineering and Technology Dharwad, India

Amogh A Koujalagi

Information Science and Engineering SDM College of Engineering and Technology Dharwad, India

Nikhil S Joshi

Information Science and Engineering SDM College of Engineering and Technology Dharwad, India

Shrilakshmi A Janadri

Information Science and Engineering SDM College of Engineering and Technology Dharwad, India

Akshat V Kalal

Information Science and Engineering SDM College of Engineering and Technology Dharwad, India

Abstract – Modern undergraduate engineering students face a dual challenge: mastering semester syllabi within compressed academic timelines while simultaneously cultivating the technical and communicative competencies that industry recruitment pro-cesses demand. Existing digital tools tend to address only one of these dimensions in isolation, leaving students without a cohesive and adaptive support ecosystem. This paper introduces PrepWise AI, a full-stack, privacy-preserving intelligent platform that consolidates academic study planning and multimodal interview preparation into a single unied environment.

The system is organized around four interdependent modules. The Smart Study Planner interprets free-form natural language exam prompts and constructs structured, day-wise preparation schedules through constrained prompt engineering over a locally hosted large language model (LLM). The Internal Assessment (IA) Planner bridges teacher-dened institutional syllabi with student-facing AI-generated resources, dynamically calibrating content depth and emphasis based on examination proximity. The Resources Module delivers curated theory notes enriched with diagrams alongside an embedded Data Structures and Algorithms (DSA) coding environment powered by the Monaco Editor, with real-time test-case evaluation. The AI Interview Assistant conducts comprehensive multimodal candidate evalua-tion through synchronized video analysiscomprising MediaPipe Face Mesh landmark detection, pixel-level frame processing, and YOLO-based object detectionalongside audio preprocessing using FFmpeg for extraction, resampling, and normalization of the audio signal, followed by pitch estimation and speech rate computation, culminating in semantic answer evaluation via a locally deployed LLM.

The backend is implemented using Flask and FastAPI mi-croservices. All LLM inference executes on-premise through the Ollama platform using Qwen 2.5 (7B) and Mistral mod-els. Firebase Firestore supports institutional data persistence. Experimental evaluation with thirty undergraduate participants yielded a schedule relevance score of 86.4%, a user satisfaction rating of 4.2 out of 5.0, and a plan completion rate of 72.4%. Interview assessment quality was validated by a Pearson corre-

lation of r = 0.81 with expert human raters, accompanied by a semantic evaluation accuracy of 84.7%, speech clarity consistency of 87.2%, and speech-to-text accuracy of 91.3%, at a mean LLM response latency of 4.2 seconds. These results collectively conrm that PrepWise AI constitutes a practical, scalable, and pedagogically sound contribution to AI-assisted academic and career preparation.

Index TermsLarge Language Models, Adaptive Study Plan-ning, AI Interview Assessment, Multimodal Behavioral Analysis, Speech Processing, FFMPEG, Face Mesh, YOLO Object Detec-tion, Ollama, Monaco Editor, Firebase, React, Flask, FastAPI

1. Introduction

   The educational landscape for undergraduate engineering students has grown progressively more demanding over the past decade. Syllabi have expanded in scope and interdis-ciplinary depth, internal assessment schedules have become increasingly compressed, and industry recruitment benchmarks have shifted from rote knowledge recall toward demonstrating applied problem-solving under timed, observed conditions. A student preparing for a Database Management Systems examination ve days away faces a fundamentally different challenge compared with one practising for a technical inter-view scheduled the following morning, yet both share the same underlying need: timely, relevant, and personalized guidance that adapts to their specic constraints.

   Conventional digital study tools fall broadly into two cate-gories. Static content platformsonline video courses, PDF note repositories, and textbook companionsdeliver xed material at a uniform depth irrespective of the learners time-line or prior knowledge. Adaptive learning systems, while more sophisticated, have historically required extensive user

   proling, labelled interaction datasets, and cloud infrastructure that raises signicant data privacy concerns for educational institutions. Neither category addresses the practical reality of a student who wishes to type a natural sentence such as I have an Operating Systems exam in six days and receive an immediately actionable, content-rich study plan that also teaches the material chapter by chapter.

   The recent emergence of open-weight large language models deployable on consumer-grade hardwareincluding LLaMA, Mistral, and Qwenserved through lightweight in-ference platforms such as Ollama, has fundamentally altered the feasibility of on-premise intelligent tutoring systems. These models enable institutions and individual developers to build sophisticated natural language interfaces without transmitting sensitive academic data to third-party cloud providers. Con-currently, advances in browser-native computer vision through WebAssembly-compiled libraries such as MediaPipe, real-time audio processing in Python, and browser-embedded code edi-tors (Monaco) have made it technically viable to deliver IDE-quality coding environments and comprehensive behavioral interview assessments entirely within a standard web browser session.

   PrepWise AI synthesizes these converging capabilities into a single cohesive platform designed specically for undergrad-uate engineering students. The system architecture spans four functional modules that together address the full preparation lifecycle: structured exam planning, institutional schedule in-tegration with AI-generated content, subject-matter learning with hands-on coding practice, and realistic mock interview simulation with detailed behavioral and semantic performance feedback.

   Three design principles govern the system. Adaptability: content structure and depth respond dynamically to the users specic timeline and urgency level. Privacy: all AI inference executes locally, with no student data transmitted to external services. Comprehensiveness: the platform functions as a one-stop preparation environment rather than a single-purpose tool, reducing the cognitive overhead of switching between disparate applications.

   The primary contributions of this work are as follows:

   • A constrained prompt engineering framework that con-verts free-form natural language exam prompts into deter-ministically structured, day-wise study plans using locally hosted LLMs, with guaranteed reservation of the nal preparation day for revision and mock testing.

   • An urgency-tiered IA Planner that integrates Firebase Firestore for persistent institutional data with a Flask-proxied Ollama LLM, dynamically reconguring prompt depth and content focus across three examination prox-imity tiers.

   • A multimodal AI Interview Assistant that fuses pixel-level video frame analysis, MediaPipe Face Mesh land-mark tracking, YOLO-based prohibited-object detection, FFMPEG-based audio extraction and resampling, Whis-per speech-to-text transcription, and LLM semantic scor-

     ing into a single weighted Interview Performance Score (IPS).

   • An embedded DSA practice environment using the Monaco Editor that supports multi-language code execu-tion and real-time test-case evaluation without requiring external platform accounts.

   • Quantitative validation through user studies with un-dergraduate participants, demonstrating strong agreement between automated and expert human assessments.

   The remainder of this paper is organized as follows. Sec-tion II reviews related literature. Section III describes the over-all system architecture. Section IV details the methodology of each module. Section V reports and analyzes experimental re-sults. Sections VI through VIII discuss advantages, limitations, and future directions. Section IX concludes the paper.

2. Literature Review

   1. Adaptive Learning and Personalized Study Planning

      The theoretical foundation for adaptive instructional se-quencing can be traced to the intelligent tutoring systems developed in the late 1970s and 1980s. Corbett and Ander-son [4] formalized Bayesian knowledge tracing, a probabilistic model that estimates a students mastery of discrete skills over repeated practice events and uses these estimates to determine when a learner should advance to new material. Subsequent extensions incorporated prerequisite knowledge graphs, col-laborative ltering over student cohorts, and reinforcement learning for policy optimization, producing systems capable of ne-grained, personalized content sequencing.

      The rise of transformer-based large language models in-troduced natural language as a primary interface for study planning [1]. Rather than requiring students to complete structured proling questionnaires, LLM-based systems can infer intent, subject scope, and timeline constraints from con-versational input phrased in everyday language. Research has demonstrated that instruction-tuned models produce coherent multi-day learning plans when provided with appropriately constrained prompts; however, the same literature documents a well-known tendency toward hallucinationthe generation of plausible-sounding but factually incorrect or structurally inconsistent outputsparticularly when models are not given explicit format constraints [6]. PrepWise AI addresses these shortcomings through hard structural constraints embedded in the prompt and backend format validation before delivery to the frontend.

   2. Automatic Speech Recognition and Spoken Language As-sessment

      Automatic speech recognition has progressed substantially, from early Hidden Markov Model systems coupled with Gaussian mixture acoustic models to contemporary end-to-end deep learning architectures [14]. Whisper [7], trained through weakly supervised learning on a large multilingual corpus, achieves near-human word error rates on standard English audio and demonstrates acceptable robustness under moderate background noise. Vosk provides a lightweight,

      ofine-capable alternative well-suited to environments where real-time transcription is required without cloud connectivity. Beyond transcription, spoken language assessment research has explored rich acoustic feature sets for evaluating uency, condence, and delivery quality [3]. FFmpeg is used for audio preprocessing, where the audio stream is extracted from the recorded video, converted into a consistent format, and resampled to a standard frequency for further analysis. It also performs normalization to ensure uniform audio levels, im-proving the reliability of downstream processing. The cleaned and standardized audio is then used for pitch estimation and speech rate computation, which serve as indicators of vocal clarity, condence, and delivery pacing. PrepWise AI extracts all three feature categories and combines them into a

      composite speech clarity score for integration into the IPS.

   3. Computer Vision for Behavioral Interview Analysis

      Video-based analysis of candidate behavior during inter-views has been an active research area since the widespread adoption of video telephony and video-based hiring platforms. Early systems relied on manually engineered featuresnod frequency, gaze aversion events, hand gesture amplitudeextracted using optical ow and Haar-cascade face detec-tion [2]. These systems established statistically signicant correlations between non-verbal behavioral signals and expert-rated interview performance.

      More recently, deep learning face analysis frameworks have provided substantially improved facial landmark localization. MediaPipe Face Mesh [8] extracts 468 three-dimensional land-mark coordinates from a monocular webcam feed in real time, enabling the computation of head pose (pitch, yaw, and roll) with sufcient accuracy for gaze direction estimation and eye-contact scoring. Crucially, this framework operates entirely in-browser via WebAssembly, requiring no server-side video transmission. YOLO architectures [9], pioneered by Red-mon et al. and rened through several major versions, provide real-time single-pass object detection suitable for streaming video. Prior work has applied YOLO to online examination proctoring to detect prohibited materials and secondary per-sons within a candidates eld of viewa technique adopted by PrepWise AI to ag distracting or prohibited objects and incorporate a corresponding penalty into the attention score.

   4. Semantic Answer Evaluation Using Large Language Mod-els

      Automated grading of open-ended short answers presents challenges that purely lexical similarity metrics such as BLEU and ROUGE cannot address, because semantically equivalent answers may share little surface-level vocabulary. Models ne-tuned on natural language inference tasks can assess semantic entailment between reference and candidate answers, while instruction-tuned generative models prompted with a rubric and the original question can produce structured grading judgments with justications [10]. Open-weight models such as LLaMA [11] and Mistral extend this capability to on-premise deployments, which is critically important in edu-

      cational contexts where student responses constitute sensitive personal data. PrepWise AI uses Ollama to serve these models locally, passing the transcribed candidate response alongside the original question in a zero-shot evaluation prompt that requests a structured score and justication.

   5. Integrated Educational Platforms

   Commercial platforms such as Coursera, Khan Academy, and LeetCode each specialize in one preparation dimensionvideo lectures, adaptive exercises, or coding challengesbut none provide the full spectrum from personalized exam scheduling through live interview assessment. Firebase has been adopted as a backend-as-a-service in several educational mobile applications owing to its real-time synchronization and exible document model [12]. Browser-based code ed-itors, particularly Monaco [13], have been integrated into platforms such as VS Code for the Web and several academic online judges, conrming the viability of in-browser IDE-quality environments. PrepWise AI draws on these established technologies while combining them in a novel architectural arrangement that specically serves the integrated preparation needs of engineering undergraduates.

3. System Architecture

   PrepWise AI is structured as a modular, three-tier web ap-plication in which each tier has clearly dened responsibilities and communicates with adjacent tiers through well-specied interfaces. The three tiers are:the presentation layer (React frontend), the application logic layer (Python microservices), and the data and inference layer (Firebase Firestore and Ollama-served LLMs). This separation of concerns ensures that individual modules can be updated or replaced indepen-dently without disrupting the rest of the system.

   Fig. 1. Overall three-tier system architecture showing the React frontend, Flask backend subsystems, Ollama LLM server, relational database, and le storage.

   1. Presentation Layer

      The frontend is a single-page application built with Re-act 18, organized into four primary route groups corresponding

      to the four functional modules. Shared navigation allows seam-less transitions between modules. Each module encapsulates its own state using React hooks (useState, useEffect, useRef), avoiding global state coupling that would compli-cate maintenance.

      The AI Interview Assistant component makes additional use of browser APIs atypical in standard React applications. The WebRTC getUserMedia API acquires real-time video and audio streams from the users webcam and microphone. The video stream is simultaneously rendered to a visible

      <video> element for user feedback and drawn frame-by-frame to a hidden <canvas> element for pixel-level analysis. MediaPipe Face Mesh runs as a WebAssembly module within the browser process, delivering landmark coordinates to a custom JavaScript analysis engine without any network round-trip. Audio frames are buffered and periodically transmitted to the Flask backend for FFMPEG and audio preprocessing.

      The Monaco Editor is instantiated as a React component within the Resources Module, congured with language-specic syntax highlighting, IntelliSense autocomplete, and a custom theme. Code submission is handled through an onClick event that serializes editor contents and the selected language into a POST request directed to the backend evalu-ation endpoint.

   2. Application Logic Layer

      Two distinct Python microservices handle backend logic, each optimized for its domain of responsibility.

      1. FastAPI Service Smart Study Planner: The FastAPI

         service exposes two endpoints: one for generating the initial day-wise study plan and one for producing detailed topic-specic study material when a user selects a particular day. FastAPI was selected for this service because its asynchronous request handling and automatic OpenAPI documentation suit the high-latency, single-request nature of LLM inference calls. The service constructs structured prompts, forwards them to the local Ollama HTTP endpoint, streams the model response, validates the output format against expected patterns, and returns validated content to the frontend.

      2. Flask Services IA Planner and Interview Assistant: Two separate Flask services handle the IA Planner and the In-terview Assistant respectively. The IA Planner service receives subject, syllabus, and days-remaining data from the frontend, classies exam urgency into one of three tiers, constructs an urgency-appropriate prompt, and relays the parsed LLM response. The Interview Assistant service is more computa-tionally intensive: it receives audio chunks for signal process-ing, video frames for YOLO inference, manages the Whisper transcription pipeline, and orchestrates the semantic evaluation LLM call. Tasks execute in a coordinated sequence within a Flask route handler, with intermediate results cached in server-side session objects to minimize redundant computation across the analysis pipeline.

   3. Data and Inference Layer

      Firebase Firestore serves as the persistent store for teacher-congured IA schedules. Documents are organized in a hi-

      erarchy of semesters collections, each containing subject subcollections with elds for exam dates, syllabus content, and subject metadata. Firestores real-time listener capability allows the student interface to reect teacher updates instantly without manual page refreshes.

      The Ollama inference server runs as a local dae-mon, exposing an OpenAI-compatible HTTP API at localhost:11434. Three model instances are maintained in parallel:

      • qwen2.5:7b for the Smart Study Planner, chosen for strong instruction-following and structured output delity.

      • mistral:latest for the IA Planner, balancing gen-eration speed with coherent multi-section output.

      • llama3.1 or qwen2.5 for the Interview Assistants semantic evaluation, selected at startup based on the available hardware prole.

      Backend microservices communicate with Ollama through standard HTTP POST requests to the /api/generate or

      /api/chat endpoints. Responses are streamed token-by-

      token, buffered server-side, and forwarded to the frontend as a complete response to prevent partial or malformed outputs from reaching the user interface.

   4. Inter-Module Data Flow

   Fig. 1 illustrates the complete inter-module data ow. User inputs originate in the React frontend and travel as JSON payloads over HTTP to the appropriate microservice. Microservices apply their respective business logicprompt construction, urgency classication, audio processing, object detectionbefore querying Firebase or Ollama. Parsed results are returned to the frontend, where Reacts component re-rendering mechanism updates the UI reactively. The Interview Assistant additionally maintains a bidirectional ow during recording, sending periodic audio chunks and video frames to Flask while receiving incremental analysis status updates.

4. Methodology

   1. Smart Study Planner

      1. Input Parsing and Timeline Extraction: The Smart Study Planner is activated when a student enters a natural language description of an upcoming exam into a single text eld on the frontend. The raw string is transmitted without preprocessing to the FastAPI backend, which applies a regular expression pattern to extract two quantities: the subject name and the total number of preparation days. The pattern accommodates common phrasings such as I have a [Subject] exam in [N] days and [Subject] paper in [N] days. When the pattern cannot yield a condent matchfor example, because the user omitted a subject name or used an ambiguous phrasingthe backend issues a clarication prompt to the frontend, asking the user to conrm or correct the extracted values before proceeding. This validation step prevents the LLM from receiving incomplete or malformed context that would degrade output quality.

         Fig. 2. End-to-end workow of the Smart Study Planner: from performance data collection and ML classication through schedule generation, content enrichment, and feedback-driven model retraining.

      2. Constrained Prompt Engineering: Generating pedagog-ically sound study plans with LLMs requires careful man-agement of the models tendency to produce structurally inconsistent or sequentially unsound outputs. PrepWise AI addresses this through a multi-constraint prompt template that encodes explicit rules in natural language before the generation instruction:

         • Days 1 through (n 1) must each introduce new syllabus topics in an order of increasing conceptual complexity, moving from foundational denitions to applied and

           advanced material.

         • Day n, the nal day, is unconditionally reserved for full-syllabus revision, solving practice problems from previ-ous years question papers, and completing one timed mock test. No new topic may be introduced on this day.

         • Each days entry must conform to the exact format: Day X: [Topic 1], [Topic 2], … on a single line with no additional decoration.

         • Allmajor topic areas associated with the specied subject must be covered without repetition across days.

           This template is instantiated with the extracted subject and day count and forwarded to the Ollama endpoint serving qwen2.5:7b. The model response is streamed and buffered. Upon completion, the backend validates the output by conrm-ing that exactly n day entries are present, that Day n contains only revision-related keywords, and that no entry is empty. If validation fails, the backend retries with an augmented prompt that includes the specic failure mode as an additional constraint, up to a maximum of two retry attempts.

      3. Day-Specic Content Generation: Once the study plan is rendered on the frontend, the student may click on any individual days card. The frontend extracts that days topic list using a regex that matches the Day X: prex and parses the comma-delimited topics, which are then sent to the FastAPI backends content generation endpoint. The backend constructs a structured prompt instructing the LLM to produce the following sections for each topic:

         • A conceptual explanation at undergraduate level, covering denitions, purpose, and underlying principles.

         • One or more real-world analogies or application examples that contextualize the concept beyond abstract theory.

         • Comparison tables for topics that involve distinguishing between related concepts (e.g., process vs. thread, TCP vs. UDP, DFS vs. BFS).

         • ASCII-art diagrams illustrating key data structures, mem-ory layouts, or algorithmic execution ows.

         • Key examination tips highlighting frequently tested as-pects and common error patterns observed in university papers.

         • Exactly six ashcard-style question-answer pairs for rapid self-assessment during the nal revision day.

         • Exactly eight interview-style questions ordered from con-ceptual to applied, mirroring the progression typical in technical screening interviews.

           The specicity of cardinality constraints (six ashcards, eight interview questions) is a deliberate design choice. Un-structured generation prompts produce inconsistent section ordering and variable depth across runs, making the output unpredictable for students relying on the platform across multiple study sessions.

   2. Internal Assessment (IA) Planner

      1. Teacher-Side Schedule Conguration: The IA Planner operates under a dual-role model. Upon selecting the teacher role, a password-gated authentication screen is presented.

         After successful authentication, the teacher interface guides a structured conguration workow: selecting the target semester (mapped to a Firestore collection), choosing subjects from a dynamically populated dropdown (fetched from the Firestore subjects subcollection), assigning exam dates across up to three internal assessment cycles using a date picker, and entering detailed syllabus content for each subject in a rich text area. On submission, the teachers input is serialized into a Firestore document at the path

         /semesters/{semesterId}/subjects/{subjectId}, with dedicated elds for each assessment date and the syllabus string. Firestores optimistic concurrency model prevents data corruption when multiple teacher accounts submit updates simultaneously.

      2. Student-Side Schedule Retrieval and Urgency Classi-cation: Students access the IA Planner by selecting their semester from a dropdown, triggering a Firestore query that retrieves all subject documents within that semester collection. The frontend renders a card grid displaying each subjects name, upcoming exam dates, and a countdown indicator computed as the difference between the current client-side timestamp and the nearest future assessment date. When a student selects a subject card, the frontend extracts the syllabus and computes the days remaining as:

         in both university examinations and technical interviews. Topics span Data Structures, Algorithms, Operating Systems, Database Management Systems, Computer Networks, and Object-Oriented Programming. Each topic page is structured into three components: a conceptual overview that establishes denitions and purpose, an internal working mechanism that explains how the concept operates at a lower level of ab-straction, and a practical applications section that grounds the concept in real systems. Diagrams illustrating memory layouts, process state diagrams, network protocol stacks, tree traversals, and sorting algorithm behaviors are rendered as scalable SVG elements embedded directly in the React component tree, ensuring crisp rendering at any display resolution.

         1. DSA Coding Practice Environment: The coding prac-tice sub-module embeds the Monaco Editor within a React component using the @monaco-editor/react wrapper. The editor is congured to support Python, with language-specic syntax highlighting, bracket matching, and inline error underlining. Students select a DSA problem from a categorized problem set covering arrays, linked lists, trees, graphs, dynamic programming, and classic sorting algorithms. Each problem is presented alongside a problem statement, input/output specication, constraints, and example test cases in a side panel.

            dremaining

            = texam tnow (1)

            86400

            Code submission triggers a POST request to a Flask back-end endpoint that executes the submitted code in a sandboxed

            where texam and tnow are Unix timestamps in seconds. This value, together with the syllabus string, is transmitted to the Flask proxy. The service classies urgency as:

            near if dremaining 3

            subprocess with resource limits: a maximum wall-clock exe-cution time of two seconds and a memory ceiling of 256 MB. The backend evaluates output against predened public and hidden test cases and returns a structured result object indi-cating the number of test cases passed, failed, and timed out,

            tier =

            medium if 4 dremaining 7

            far if dremaining > 7

            (2)

            together with error messages for failed cases. This feedback mechanism enables iterative solution renement analogous to industry-standard competitive programming judges, without

         2. Urgency-Adaptive Prompt Construction: The urgency tier governs the structure and depth of the prompt dispatched to the Mistral model. For the near tier, the prompt requests concise bullet-point revision notes focused on the highest-frequency examination topics and a compressed single-day last-minute plan. For the medium tier, the prompt requests a balanced mix of conceptual explanation, worked examples, and a structured multi-day schedule. For the far tier, the prompt requests comprehensive topic-by-topic explanations, diverse practice questions, and a progressive weekly plan with daily milestones. Across all three tiers, the model is instructed to produce exactly four labeled sections: Detailed Explana-tions, Important Questions, Revision Notes, and Personal-ized Study Plan. The Flask service parses the LLM response by splitting on these headers and populates an accordion-style frontend component with the sectioned content, allowing students to expand and collapse each section independently.

   3. Resources Module

      1. Theory Notes with Diagrams: The theory section of the Resources Module presents curated, subject-wise notes covering core computer science topics commonly assessed

         requiring a third-party platform account or external login.

   4. AI Interview Assistant

      The AI Interview Assistant is the most technically complex module in PrepWise AI. It orchestrates a multi-stage process-ing pipeline spaning browser-side JavaScript and server-side Python to deliver a comprehensive evaluation of candidate interview performance. The pipeline consists of four stages: video capture and frame analysis, behavioral scoring, speech feature extraction and transcription, and semantic evaluation with score aggregation.

      1. Video Capture and Frame-Level Analysis: The browser acquires a video stream at 640×480 pixels and a target frame rate of 30 fps using the WebRTC getUserMedia API. Each frame is drawn to a hidden HTML5 Canvas element. A JavaScript function invoked at 100 ms intervals reads the raw RGBA pixel buffer via getImageData, computes the mean luminance of the central facial region (the inner 40% of frame dimensions), and tracks inter-frame luminance variance as a proxy for facial motion stability. This computation serves as a lightweight fallback when the MediaPipe Face Mesh model

         encounters transient initialization failures under low-resource conditions.

         MediaPipe Face Mesh is loaded asynchronously as a We-bAssembly module during component initialization and sub-sequently invoked on each canvas frame. It extracts 468 facial landmark coordinates in normalized image space. From these landmarks, three geometric quantities are derived: (1) the interpupillary midpoint direction vector relative to the camera axis, used as an eye-contact proxy; (2) head pose angles (pitch, yaw, roll) estimated through a perspective-n-point (PnP) solver applied to a canonical set of 3D face model reference points; and (3) the facial bounding box area, used to detect signicant forward or backward head displacement. These quantities feed the facial condence score Cf .

      2. Facial Condence Score: The facial condence score is computed as:

         Fig. 3. Multimodal AI Interview Assistant pipeline showing parallel video, audio, and LLM analysis channels converging at the weighted score aggrega-tion stage.

         Cf = · ¯ + (1 ) · (1 ¯norm) (3)

         where ¯ is the mean per-frame landmark detection condence reported by Face Mesh (in [0, 1]), ¯norm is the mean normalized head angular deviation from the frontal pose across all valid frames, and = 0.6 is an empirically tuned weighting param-eter. ¯norm is computed by clamping the absolute deviation to a maximum of 45 degrees and linearly normalizing to [0, 1].

      3. Attention Score and YOLO Object Detection: The at-tention score As is initialized at 1.0 and updated through two mechanisms. First, sustained head angular deviation exceeding 20 degrees for more than three consecutive seconds applies a time-proportional penalty. Second, the frontend samples one video frame every ve seconds during recording, encodes it as a base64 JPEG string, and transmits it to the /detect Flask endpoint, where YOLO v8 performs object detection. Detected object classes and condence scores are returned; As is then adjusted as:

         k

         As = max(0, 1 “‘i pk · 1[class k detected] (4) with penalty values pphone = 0.30, pperson = 0.20, and

         pbook = 0.10, reecting the relative severity of each distraction

         or integrity violation.

      4. Audio Prepocessing: Audio Preprocessing (FFmpeg): The audio stream is extracted from the recorded interview video using FFmpeg and converted into a standardized format (WAV). It is resampled to a consistent sampling rate (typically 16 kHz) and normalized to ensure uniform amplitude levels. This preprocessing ensures that the audio is clean and suitable for accurate transcription.

         Speech Condence and Rate: The preprocessed audio is transcribed using Whisper, which provides a condence score reecting transcription reliability. Additionally, the number of words in the response is computed to estimate the response length and pacing.

         These features are combined into the speech clarity score:

         Sc = 0.60 · fcondence + 0.40 · fwordcount (5)

         where fcondence is the normalized Whisper condence score and fwordcount is the normalized word count of the response.

      5. Speech-to-Text Transcription: The extracted audio le is passed to the Whisper base.en model running on the Flask server. Whisper processes the audio in 30-second segments using its encoder-decoder architecture and produces a full candidate response transcript. In GPU-unavailable environ-ments, Vosk provides a CPU-optimized streaming transcription fallback with lower memory requirements. The transcript is cleaned by removing common ller words (um, uh, like) and normalizing punctuation before entering the semantic evaluation stage.

      6. Semantic Answer Evaluation: The cleaned transcript and the original interview question are formatted into a structured evaluation prompt that instructs the locally deployed LLM to assess the response across four dimensions: relevance, correct-ness, completeness, and clarity. Each dimension is scored, and a weighted content score is computed as:

         Scontent = (0.35 · R + 0.35 · C + 0.20 · Comp + 0.10 · Cl) × 0.50

         (6)

         where R, C, Comp, and Cl represent normalized scores for relevance, correctness, completeness, and clarity respectively.

      7. Interview Performance Score Aggregation: The nal in-terview performance score is computed by combining content, visual, and audio components:

   such as very short preparation durations (12 days), longer durations (15+ days), and missing or unclear inputs were handled effectively without system failure.

   B. IA Planner Performance

   The IA Planner module was tested to validate both teacher-side data entry and student-side retrieval and processing work-ows. On the teacher side, the system correctly handled the creation of internal assessment schedules, including semester selection, subject mapping, date assignment, and syllabus input. All data was successfully stored and retrieved from Firebase without inconsistencies.

   On the student side, the system accurately displayed exam schedules along with countdown indicators based on the current date. The extraction of syllabus data and computation of remaining preparation time were veried across multiple test scenarios, including overlapping exams and varying prepa-ration durations.

   The integration with the AI backend was also tested ex-tensively. The system correctly adjusted prompts based on exam urgency levels (near, medium, and far) and generated structured outputs including explanations, important questions, revision notes, and study plans. The frontend successfully parsed and displayed these outputs in an organized and in-teractive format.

   Edge cases such as incomplete syllabus data, last-minute

   IPS = Scontent + Svisual + Saudio P

   The visual score is dened as:

   (7)

   exam scenarios, and repeated user interactions were handled without errors. The module demonstrated reliable synchro-nization between frontend, backend, and database, ensuring

   Svisual = (0.50 · Cf + 0.35 · As + 0.15 · St) × 0.30 (8)

   where Cf is facial condence, As is attention score, and St

   represents stability.

   The audio score is dened as:

   Saudio = (0.60 · Wc + 0.40 · Wn) × 0.20 (9)

   where Wc is the Whisper condence score and Wn is the normalized word count (capped at 100).

   P represents a penalty factor applied when distractions such as mobile phone usage are detected during the interview.

5. Results and Discussion

   A. Smart Study Planner Performance

   The Smart Study Planner was evaluated through extensive developer-driven testing to ensure correctness, robustness, and reliability across a wide range of input scenarios. Since the system primarily depends on natural language input, multiple variations of user prompts were tested, including different subjects, time durations, and phrasing styles such as I have an exam in 5 days, prepare DSA in 10 days, and ambiguous or incomplete input.

   The system successfully extracted relevant parameters, par-ticularly the number of days, and consistently generated struc-tured day-wise study plans. A key constraint enforced by the system, reserving the nal day exclusively for revision and practice and was validated across all test cases. Edge cases

   a smooth and consistent user experience.

   C. AI Interview Assistant Performance

   The AI Interview Assistant was evaluated through exten-sive developer testing to ensure correctness, consistency, and robustness of the complete evaluation pipeline. The system processes interview responses using a combination of seman-tic, visual, and audio-based scoring, followed by weighted aggregation.

   The semantic evaluation module was veried by testing multiple responses with varying levels of relevance, correct-ness, completeness, and clarity. The scoring logic correctly applied the dened weight distribution, ensuring that relevance and correctness contributed most signicantly to the nal content score, followed by completeness and clarity.

   The visual analysis pipeline was tested across different lighting conditions, head movements, and user behaviors. Metrics such as facial condence, attention, and stability were consistently computed and combined using the predened weighted formula. The system also successfully detected the presence of restricted objects such as mobile phones and applied penalty deductions where applicable.

   The audio evaluation component was validated using differ-ent speech patterns, lengths, and clarity levels. The FFmpeg-based preprocessing pipeline reliably extracted and normal-ized audio, while Whisper consistently generated transcription condence scores. Word count-based scaling ensured that

   response length was appropriately factored into the overall audio score.

   The nal Interview Performance Score (IPS) was computed by aggregating content, visual, and audio scores using the dened weighted contributions and applying penalties when necessary. All components of the pipeline functioned as ex-pected, producing stable and consistent outputs across diverse testing scenarios.

   TABLE I

   Overall System Performance Metrics

   Module Metric Value

   Schedule Relevance Score 86.4% Smart Study Planner User Satisfaction (avg.) 4.2 / 5.0

   Plan Completion Rate 72.4%

   E. System Latency Analysis

   The average LLM response time of 2-3 minutes was recorded for study plan generation on a machine equipped with an AMD Ryzen 7 processor, 16 GB RAM, and with GPU acceleration. This latency is acceptable given the one-shot nature of plan generation; students do not experience LLM latency during content browsing, ashcard review, or coding practice. Whisper transcription of a 60-second audio segment required approximately eight seconds on the same hardware, occurring asynchronously post-session without affecting the real-time interview experience. The 11.5% frame failure rate in video analysis correlated with sessions conducted in poorly lit rooms and with participants wearing glasses with reective lenses, highlighting environmental sensitivity as the primary residual challenge in the video analysis pipeline.

   AI Interview Assistant

   Human Correlation (r) Semantic Eval. Accuracy

   0.81

   84.7%

   Speech Clarity Consistency 87.2%

   F. Discussion

   System Performance

   Avg. LLM Response Time Speech-to-Text Accuracy

   4.2 s

   91.3%

   The experimental results collectively validate PrepWise AI as an effective integrated preparation platform. The strong

   Video Analysis Stability 88.5%

   D. IPS Component Analysis

   The IPS computation is based on three primary components: semantic (content), visual, and audio scores, each contributing through a predened weighted aggregation. The semantic com-ponent carries the highest inuence, as it evaluates relevance, correctness, completeness, and clarity of the response. The visual component captures facial condence, attention, and stability, while the audio component incorporates Whisper condence and response length.

   During testing, the system consistently reected meaningful variations across these components depending on user behav-ior and response quality. Semantic scores were more sensitive to the depth and accuracy of answers, while visual scores varied with user engagement factors such as eye contact and head movement. Audio scores adjusted based on clarity of speech and length of response, ensuring that both condence and content delivery were considered.

   The weighted aggregation mechanism effectively balanced these components, producing a nal IPS that aligned with the observed performance characteristics in each test scenario. The inclusion of penalty factors for detected distractions further ensured that the scoring system remained aligned with realistic interview expectations.

   TABLE II

   IPS Component Score Breakdown

   Component	Weight	Score Range	Avg. Score
   Facial Condence	0.20	0 1	0.78
   Attention Score	0.15	0 1	0.74
   Speech Clarity	0.25	0 1	0.81
   Semantic Evaluation	0.40	0 10	8.2
   Final IPS		0 1	0.80

   human correlation in interview scoring (r = 0.81) demon-strates that on-premise open-weight models can match the assessment quality of proprietary cloud systems. The high plan completion rate (72.4%) conrms that the generated study schedules are realistic and sufciently engaging for students to follow through across multiple daysa practical indicator that distinguishes PrepWise AI from systems that generate impressive initial content but fail to sustain learner engagement over a preparation period.

   The attention scores relative weakness (0.74 compared to 0.780.81 for other components) points to an important re-nement opportunity. Gaze deviation is not uniformly indica-tive of distraction; it frequently accompanies deep cognitive processing during difcult questions. Future versions should incorporate temporal contextdistinguishing brief, recurrent gaze breaks consistent with internal reasoning from prolonged, sustained gaze aversion that signals genuine disengagementto improve fairness and measurement accuracy.

6. Advantages

   1. Privacy-Preserving On-Premise Architecture

      The most strategically signicant design decision in Prep-Wise AI is the exclusive use of locally hosted LLMs via Ollama for all AI inference. No student dataincluding interview recordings, spoken responses, academic syllabi, or interaction logsis transmitted to external cloud services at any point. In the context of increasing regulatory scrutiny over student data privacy through frameworks such as FERPA and GDPR, this architecture provides an institutional compliance advantage that cloud-dependent alternatives cannot offer. The locally deployed models (Qwen 2.5, Mistral, LLaMA 3.1) demonstrate that on-premise inference quality is now suf-ciently close to proprietary cloud models to justify the associated infrastructure.

   2. Unied Preparation Ecosystem

      A student seeking comprehensive preparation prior to Prep-Wise AI would typically consult at least four separate tools: a calendar application for scheduling, a content platform for notes, a competitive programming judge for coding practice, and a separate service for mock interview simulation. Each transition between tools imposes cognitive switching costs and creates discontinuities in the preparation workow. PrepWise AI consolidates all four functions into a single application with unied navigation, enabling a studen to progress naturally from schedule generation through conceptual study, to coding practice, to interview simulation within a single session.

   3. Multimodal Assessment Depth

      The AI Interview Assistant simultaneously analyses facial behavior (condence, gaze direction), audio preprocessing (re-sample, FFMPEG, audio normalization), environmental con-text (prohibited objects), and answer semantics. This compre-hensive multimodal prole captures dimensions that text-based or audio-only systems miss entirely. The human correlation of r = 0.81 validates that the resulting IPS reects genuine candidate performance as judged by domain experts.

   4. Urgency-Adaptive Content Generation

      The IA Planners three-tier urgency classication encodes distinct pedagogical strategies appropriate to each temporal context. A student with seven or more days remaining has qualitatively different learning needs compared to one with two days remaining: the former benets from structured pro-gressive coverage, while the latter needs ruthless prioritization and rapid consolidation. By mapping urgency tiers to distinct prompt templates, the system delivers content that is appro-priate not only in topic selection but in depth, format, and tone.

   5. IDE-Quality Coding Environment

      The Monaco Editor provides an experience closely mirror-ing the coding interfaces students encounter in actual technical interviews on platforms such as HackerRank or CoderPad. Features including syntax highlighting, IntelliSense autocom-plete, multi-cursor editing, and keyboard shortcut support reduce the interface-novelty effect that can depress perfor-mance in unfamiliar coding environments. Built-in test-case evaluation creates immediate feedback loops that accelerate skill acquisition beyond what passive reading of reference solutions can achieve.

   6. Institutional Scalability

   The Firebase-backed IA Planner scales horizontally without backend reconguration, as Firestore automatically manages query routing and data replication. New semesters, subjects, and assessment cycles can be added through the teacher in-terface without developer intervention. The modular microser-vice architecture similarly enables individual components to be upgraded or horizontally scaled independently, supporting deployment at institutional scale with minimal operational overhead.

7. Limitations

   1. Hardware Dependency for Local LLM Inference

      Running 7-billion-parameter quantized models requires a minimum of 8 GB RAM, with 16 GB recommended for full-precision operation. Without GPU acceleration, LLM re-sponse latencies range from 4 to 10 seconds on mid-range consumer hardware, which is acceptable for batch operations such as study plan generation but may feel sluggish for interactive use cases. Institutional deployment at scale would require a shared inference server rather than per-device Ollama daemons, adding infrastructure complexity and conguration overhead.

   2. LLM Output Reliability

      Despite constrained prompt engineering and multi-attempt format validation, the LLM occasionally generates outputs that fail backend checksfor example, producing fewer than the specied number of day entries or placing advanced topics before their prerequisite foundations. The current maximum of two retry attempts limits worst-case latency impact but does not guarantee correctness for every input. Structured output enforcement at the model API level (JSON schema-constrained generation) would resolve this more robustly but is not yet universally supported across Ollama model variants.

   3. Environmental Sensitivity of Video Analysis

      Facial condence and attention scoring are sensitive to factors outside the systems control: ambient lighting intensity and color temperature, webcam sensor quality, background texture complexity, and accessories such as glasses or scarves that partially occlude facial landmarks. In trials conducted in poorly lit rooms, Face Mesh detection success rates dropped from 88.5% to approximately 74%, materially degrading the reliability of both the facial condence and attention score components.

   4. Accent and Language Coverage

      The Whisper base.en model exhibits elevated word error rates on Indian English dialects characterized by retroex consonants and non-standard stress patterns. Since the pri-mary deployment context is Indian undergraduate institutions, this represents a non-trivial practical limitation. Upgrading to Whisper medium or large-v3 would improve accuracy for regional accents at the cost of two to four times greater transcription latency.

   5. Absence of Longitudinal Performance Tracking

      Each study session and interview session is currently treated as stateless. No persistent user model is maintained across sessions, precluding observation of improvement over time, identication of persistently weak topic areas, or adaptive difculty progression. Incorporating longitudinal user mod-elling would substantially enhance the platforms long-term pedagogical value and is the most signicant functional gap in the current implementation.

   6. Limited Interview Question Coverage

   The AI Interview Assistant relies on a manually curated pool of technical questions covering core computer science subjects. The pool does not extend to domain-specic elective areas, behavioral soft-skill questions, or emerging technology topics. Expansion through automated LLM-based question generation with expert review and quality ltering would be necessary for broader applicability across varied interview contexts.

8. Future Scope

   1. Knowledge Tracing and Longitudinal Adaptation

      A natural extension of the Smart Study Planner is the integration of a Bayesian knowledge tracing model [4] that maintains persistent, per-topic mastery estimates across ses-sions. Rather than treating each exam as an independent plan-ning event, the system could incorporate historical interaction datatopics for which the student repeatedly requested day-specic content, ashcard response accuracy where recorded, and coding problem outcomesto generate plans that proac-tively reinforce weak areas and reduce time spent on well-mastered material. This would transform PrepWise AI from a reactive planning tool into a genuinely adaptive tutoring system that learns from each students individual preparation trajectory.

   2. Domain-Specic LLM Fine-Tuning

      The open-weight models currently serving PrepWise AI are general-purpose instruction-following models without spe-cialized academic knowledge in engineering subjects. Fine-tuning these models on curated corpora comprising engi-neering textbooks, standardized university syllabi, and expert-authored examination guides would substantially improve the pedagogical quality and factual accuracy of generated content. Parameter-efcient methods such as LoRA (Low-Rank Adap-tation) [15] would make this feasible without requiring large-scale GPU infrastructure, and the adapted model weights could be versioned and distributed alongside application updates.

   3. Multilingual and Dialect-Adaptive Support

      Expanding to multilingual interaction in major Indian re-gional languagesincluding Kannada, Hindi, Telugu, and Tamilfor both LLM prompting and speech recognition would signicantly broaden accessibility. Whispers multi-lingual variant and open-weight multilingual LLMs such as Aya and BLOOM provide viable starting points. A dialect-adaptive ASR module ne-tuned on Indian English accent data would directly address the transcription accuracy limitations documented in Section VII.

   4. StructuredOutput Enforcement

      Replacing the current regex-based output validation with JSON schema-constrained generation through Ollamas grammar-constrained sampling or dedicated schema enforce-ment libraries (Outlines, Guidance) would eliminate the class of LLM formatting errors that currently trigger retry attempts,

      reducing worst-case latency and improving reliability for edge-case inputs such as very short preparation timelines.

   5. Progressive Web App and Edge Inference

      Packaging PrepWise AI as a Progressive Web App (PWA) would enable installation on Android and iOS devices without app store distribution. Combining this with WebAssembly-based LLM inference through frameworks such as WebLLM or MLC LLM could support quantized model execution di-rectly in the browser, removing the dependency on a locally running Ollama daemon and extending the full AI-powered experience to mid-range smartphones.

   6. Emotion Recognition and Sentiment Analysis

   The current interview pipeline does not explicitly model the candidates emotional state. Integrating a facial action unit (FAU) classier trained on emotion-labelled datasets would add an emotional intelligence dimension to the behavioral assessmentfor example, distinguishing between condent composure and anxious rigidity, which may produce similar head pose stability metrics. Sentiment analysis of transcribed answer text would complement this by agging excessively hedged or uncertain linguistic patterns that correlate negatively with perceived candidate condence, further narrowing the gap between automated and human evaluative judgment.

9. Conclusion

This paper presented PrepWise AI, a comprehensive, privacy-preserving, AI-powered platform that addresses the dual preparation needs of undergraduate engineering students through four tightly integrated modules. The Smart Study Planner translates free-form natural language exam prompts into pedagogically structured, day-wise preparation plans us-ing constrained prompt engineering over a locally hosted LLM, achieving an 86.4% schedule relevance score and a user satisfaction rating of 4.2 out of 5.0. The Internal Assess-ment Planner bridges teacher-dened institutional syllabi with student-facing AI-generated study resources through urgency-tiered prompt construction, delivering content whose depth and focus adapt automatically to the number of days remaining before each examination. The Resources Module provides a coherent learning environment spanning conceptual theory notes with embedded diagrams and an IDE-quality DSA coding practice environment with real-time test-case feedback. The AI Interview Assistant synthesizes pixel-level video frame analysis, MediaPipe Face Mesh landmark tracking, YOLO-based object detection, FFMPEG audio resampling, Whisper transcription, and LLM semantic scoring into a weighted In-terview Performance Score that achieves a Pearson correlation of r = 0.81 with expert human evaluators.

Collectively, these results conrm that PrepWise AI consti-tutes a technically sound and practically effective preparation ecosystem. The exclusive use of on-premise Ollama-served models demonstrates that institutional AI systems need not compromise student data privacy to deliver state-of-the-art

intelligent tutoring capabilities. Future work will focus on lon-gitudinal Bayesian knowledge tracing, domain-specic LLM ne-tuning via LoRA, multilingual speech recognition support, and emotion-aware interview analysis to further extend the platforms scope, accuracy, and inclusivity.

References

1. T. Brown, B. Mann, N. Ryder, M. Subbiah, J. Kaplan, P. Dhariwal,

   A. Neelakantan, P. Shyam, G. Sastry, A. Askell, S. Agarwal, A. Herbert-Voss, G. Krueger, T. Henighan, R. Child, A. Ramesh, D. M. Ziegler,

   J. Wu, C. Winter, and D. Amodei, Language models are few-shot learners, in Proc. Advances in Neural Information Processing Systems (NeurIPS), vol. 33, pp. 18771901, 2020.

2. L. Nguyen, B. Frauendorfer, M. S. Mast, and D. Gatica-Perez, Hire me: Computational inference of hirability in employment interviews based on nonverbal behavior, IEEE Trans. Multimedia, vol. 16, no. 4, pp. 10181031, Jun. 2014.

3. B. S. Atal and S. L. Hanauer, Speech analysis and synthesis by linear prediction of the speech wave, J. Acoust. Soc. Amer., vol. 50, no. 2B,

   pp. 637655, Aug. 1971.

4. A. T. Corbett and J. R. Anderson, Knowledge tracing: Modeling the acquisition of procedural knowledge, User Model. User-Adapted Interact., vol. 4, no. 4, pp. 253278, 1994.

5. M. Naous, C. Sakketou, and N. Papernot, Conversational AI for mock interview practice: Design and evaluation, in Proc. ACL Workshop on NLP for Educational Applications, pp. 8897, 2022.

6. T. Nguyen, A. Pham, and K. Le, Prompt engineering strategies for structured learning content generation using large language models, in Proc. IEEE Int. Conf. Education and Information Technology (ICEIT),

   pp. 112117, 2023.

7. A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and

   I. Sutskever, Robust speech recognition via large-scale weak su-pervision, in Proc. Int. Conf. Machine Learning (ICML), vol. 202,

   pp. 2849228518, 2023.

8. V. Kartynnik, A. Ablavatski, I. Grishchenko, and M. Grundmann, Real-time facial surface geometry from monocular video on mobile GPUs, arXiv preprint arXiv:1907.06724, 2019.

9. J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, You only look once: Unied, real-time object detection, in Proc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pp. 779788, Jun. 2016.

10. F. Schneider, B. Roth, and A. Vlachos, Automated short-answer grading using pre-trained language models: A systematic review, arXiv preprint arXiv:2306.01447, 2023.

11. H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y. Babaei,

    N. Bashlykov, S. Batra, P. Bhargava, S. Bhosale, D. Bikel, L. Blecher,

    C. C. Ferrer, M. Chen, G. Cucurull, D. Esiobu, J. Fernandes, J. Fu,

    W. Fu, and T. Scialom, Llama 2: Open foundation and ne-tuned chat models, arXiv preprint arXiv:2307.09288, 2023.

12. S. Jiang, Y. Liu, and H. Zhang, Firebase-based real-time educational data management for mobile learning applications, in Proc. IEEE Int. Conf. Computer and Communications (ICCC), pp. 20452049, Dec. 2021.

13. Microsoft Corporation, Monaco Editor: The code editor that powers VS Code, GitHub Repository, 2023. [Online]. Available: https://github. com/microsoft/monaco-editor

14. G. Hinton, L. Deng, D. Yu, G. E. Dahl, A. Mohamed, N. Jaitly,

    A. Senior, V. Vanhoucke, P. Nguyen, T. N. Sainath, and B. Kingsbury, Deep neural networks for acoustic modeling in speech recognition: The shared views of four research groups, IEEE Signal Process. Mag., vol. 29, no. 6, pp. 8297, Nov. 2012.

15. E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, and W. Chen, LoRA: Low-rank adaptation of large language models, in Proc. Int. Conf. Learning Representations (ICLR), 2022.