Abroad Compass: A Multi-Agent Assistant for International Student Relocation

Abroad Compass: A Multi-Agent Assistant for International Student Relocation

Arya Mandhare

Savitribai Phule Pune University Pune, India

Ojvardhan Kale

Savitribai Phule Pune University Pune, India

Dr. Manju Pawar

Savitribai Phule Pune University Pune, India

AbstractRelocating to a foreign country for higher edu-cation is a multifaceted challenge that involves lack of legal knowledge, organizational inefciency, and emotional hurdles. Existing digital solutions are often scattered all over the internet, forcing students to manually aggregate data from fragmented sources. This paper presents Abroad Compass a comprehensive multi-agent AI based system that streamlines the relocation process. Using a LangChain-based routing architecture grounded in zero-shot intent classication router [1], the system integrates specialized agents for legal guidance using Retrieval Augmented Generation (RAG), airfare forecasting using time-series regres-sion [7], [9] with veried real time data, and social matching via SBERT embeddings [11]. Furthermore, the system employs affective computing models [17] to provide contextual emotional support. The implementation leverages a FastAPI backend and an Ionic-based front-end to provide a scalable, cross-platform solution in the form of a chat-bot application. Experimental results indicate that the modular agentic approach signicantly reduces user latency and improves the accuracy of domain-specic inquiries compared to monolithic LLM architectures, which reduces hallucinations for sensitive and real-time queries.

Index TermsMultiagent systems, Retrieval Augmented Gen-eration, SBERT, Flight Fare Forecasting, Legal NLP, Affective Computing, FastAPI, Intent Classication, Hallucination Preven-tion.

1. Introduction

   The move to a foreign country to study is often more of a logistical marathon than an academic adventure. Millions of students cross borders every year [21] and are overwhelmed by the information gap. The data available on the internet is rarely personal or integrated. A student must visit a gov-ernment portal to check visa rules, a commercial aggregator for ights, social media groups for housing, scour through the internet- fact checking every scholarship and that is only the beginning.

   This fragmentation creates a Cognitive Load [20] that can result in major errors in critical legal documents and nancial planning [20]. The psychological burden of navigating these disparate systems simultaneously frequently leads to decision fatigue and suboptimal outcomes for the relocating student.

   We developed Abroad Compass with the motive of a single source of reliable information. Rather than building a generic chat-bot, we designed a modular ecosystem where each agent is a specialist in its domain. This paper details how we lever-aged recent advances in NLP, RAG for legal grounding [4],

   SBERT for semantic matching [11], and sentiment analysis

   [17] to build a tool that functions less like a search engine and more like a relocation consultant. Inspired by multi-agent conversation frameworks [2] and generative agent architectures [3], these tasks are decomposed into specialized micro-agents orchestrated by a central cognitive router, removing the bottle-neck and hallucination hazards of monolithic Large Language Models.

2. Related Work

   Abroad Compass represents the intersection of multiple re-search domains: conversational AI, legal information retrieval, predictive analytics, semantic matching, and affective com-puting. This section reviews the most relevant prior work in each area. Conversational AI has evolved from simple intent-response pairs to complex, state-aware systems. Chatbots for academic settings have been studied for university FAQ han-dling [23] and e-learning query resolution [24]. More recently, legal advisory bots [25] and student support frameworks [26] have explored domain-specic deployments. The multi-agent conversation paradigm, formalized by AutoGen [2], enables agent specialization at scale.

   1. Legal RAG Agent

      The problem of hallucination is especially concerning in the legal eld. Research on Legal-BERT [4] shows that domain-specic pre-training is necessary to capture the sub-tlety of immigration-related documents. Retrieval-Augmented Generation (RAG) via dense passage retrieval [5] has become the benchmark approach for grounding model outputs in veried source documents [6].

   2. Predictive Analytics in Travel

      Airfare pricing is highly volatile, it depends on macroe-conomic variables, seasonal trends, and route trafc density and other factors. Prior work has explored GRU-based deep learning models [7], multi-attribute dual-stage attention mech-anisms [8], and ensemble machine learning approaches [9] for fare prediction.

   3. Semantic Matching and Recommendation

   Point-of-interest recommendation has been addressed through context-aware interpretable frameworks such as

   CAPRI [14]. Social recommendation via Graph Neural Net-works [13] enables community-aware suggestions. For room-mate matching, personality-aware systems [27] have shown methodologies beyond simple hard-lter approaches.

3. System Architecture

   Abroad Compass employs a distributed, multi-layered archi-tecture designed for scalability and low-latency communica-tion. Through integration of these layers, the system undergoes seamless transitions from unstructured user inputs to intent-specic outputs guided by its dedicated agent.

   A. Authentication Process

   The system implements user access via JSON Web Tokens (JWTs) [22] for authentication via stored credentials (e.g., user email and hashed password using bcrypt) in a SQLite database during the registration process based upon elds col-lected within the users prole, including their home country, destination, eld of study, and living preference(s). When a user logs in, the system compares the password entered against the stored. If they match, a secure token based upon the users unique ID will be generated and passed on during following requests for access to secured functionality. FastAPI will then verify the token, retrieve the users prole from the database, and utilize that prole to provide secure and customized services meeting the users needs.

   B. Presentation Layer (Frontend)

   The frontend was developed using React along with Ionic framework for providing cross-platform performance. There are four key areas in the application: authentication, chatting, roommate matching, and searching for scholarships. These pages can be accessed via Ionic React Router. In order to maintain secure connections during the user session, there is an AuthProvider, which deals with the user proles and JWTs

   D. Cognitive Core: The LangChain Router

   All inbound requests are channeled into the Agentic Router once they enter the system. The LangChain Router Agent uti-lizes a zero-shot ReAct (Reason + Act) approach [1], through which the relevant agent is selected based on the context of the request. The usage of smaller, quantized local models for intent classication reduces the expenses and latency associated with calling a larger LLM model for routing purposes.

   Algorithm 1 Zero-Shot ReAct Routing Logic

   Require: User Query Q, Conversation History H, Set of Agent Tools T = {t1, t2, …, tn}

   Ensure: Selected Tool Output O

   1: Prompt FormatContext(Q, H, T )

   2: Intent LocalLLM.Clasify(Prompt)

   3: if Intent matches tk T then

   4: ExtractedArgs ParseArguments(Q, tk.schema)

   5: O ExecuteTool(tk, ExtractedArgs)

   6: else

   7: O DefaultConversationalAgent(Q, H)

   8: end if

   9: return O

4. Detailed Agent Implementation

   Abroad Compass relies on ve specialized agents, each governed by distinct algorithmic paradigms tailored to their specic domains.

   1. Legal Guidance via RAG

      The Legal Guidance Agent answers visa and immigration queries across three jurisdictions (UK, Germany, Australia) via a three-stage RAG pipeline [6]:

      [22]. Whenever an user attempts to open some restricted pages

      • Indexing: Ofcial

        immigration PDFs

        while not being signed in, application will be redirected login page immediately. In terms of the visual component, there is use made of such common Ionic elements as cards, lists, and forms. All requests to the backend were developed using RESTful APIs provided by FastAPI with a security token added to each of them.

        C. Application Gateway

        are segmented using LangChains RecursiveCharacterTextSplitter into 1,000-character chunks with 150-character overlap. Each chunk is encoded using all-MiniLM-L6-v2 [11] and stored in a ChromaDB vector database.

      • Retrieval: A user query Q is mapped into a dense vector space E : X Rd using the same encoder. ChromaDB retrieves the top-k chunks via cosine similarity:

        The application gateway is constructed using FastAPI, the

        sim(Q, D ) = q · dk

        (1)

        backend is designed as a gateway that intelligently sends requests to different services. The application is properly con-gured to talk to the frontend and consists of seven modules, namely, authentication, ights, scholarships, roommates, legal advice (RAG), and chat. Each module acts as a standalone mi-croservice with its own logic. Therefore, when problems arise for specic AI services like ChromaDB and TinyLLaMA, the entire application keeps functioning uninterrupted. Through the utilization of microservices, the application becomes scal-able as the gateway would provide security and error handling for the entire application.

        k /lq/l/ldk/l

        The jurisdiction is inferred automatically from query context or the user prole.

        • Generation: Retrieved chunks are passed to a locally-hosted, quantized TinyLLaMA (1.1B parameters) [19] with a system prompt enforcing strict source adherence, preventing hallucination of immigration law details. Re-sponses include source citations and follow-up checklists.

          The fully ofine architecture ensures sensitive student data is never transmitted to external servers [25].

          Fig. 1. Detailed System Architecture of Abroad Compass. The gure illustrates the communication between the Ionic Frontend and the FastAPI Backend, highlighting the central role of the LangChain Routing Agent in orchestrating specialized subagents including Legal, Emotional Support, Flight Fare, Scholarship, and Location Recommenders.

   2. Flight Fare Forecasting Agent

      The Flight Fare Agent employs a hybrid forecasting ar-chitecture combining an LSTM-based time-series model [10]

      dence score c is assigned based on deviation between predicted y and live price ylive:

      0.95 || 0.15y (within range)

      for temporal price trend analysis [7], and real-time market validation via the Amadeus API.

      • Feature Engineering: An 11-dimensional input vector

        c = ag deal ylive < y 0.15y

        wait ylive > y + 0.15y

        (4)

        x is constructed with encoded origin/destination airport codes, lead time (days before departure), base route price, seasonal multiplier, booking window adjustment, departure month, departure day-of-week, booking month, booking day-of-week. The base price incorporates class multipliers and a piecewise booking window function:

        1.5 t< 7 days

        (t) =

        1.3 t< 14 days

        1.1 t< 30

        days

        (2)

        This hybrid approach leverages historical price patterns from the ML model while ensuring users receive up-to-date ight details [8].

   3. Roommate Matching Agent

      The Roommate Matching Agent matches international stu-dents based on lifestyle compatibility by performing weighted content-based ltering [27], supplemented by learned attribute

      weights and semantic re-ranking via Sentence-BERT [11]. Each user prole records the following: budget range,

      0.95 t< 120 days

      0.9 t 120 days

      • Price Prediction: A serialized scikit-learn model gen-erates a predicted price y from x. An LSTM network

        housing type, cleanliness level c [1, 5], smoking and pet

        i

        preferences, study habits, sleep schedule, personality tags, and interests. Per-attribute similarity scores si(A, B) [0, 1] are computed as follows. Cleanliness proximity uses a normalized distance:

        [7], [10] independently models temporal fares to capture seasonal trends, with then undergoes ensemble prediction, given by:

        sclean(A, B) = max

        (0, 1

        |cA cB|

        4

        (5)

        B

        y = 1 X f (x) (3)

        Interest overlap is measured via the Jaccard index:

        B b

        b=1

        sint(A, B) =

        |IA

        IB|

        (6)

      • Live Market Reconciliation: The predicted range is cross-validated against live offers retrieved via the Amadeus Flight Offers Search API (OAutp). A con-

   |IA IB|

   Attribute weights w = [w1,…, w6] are learned via soft-max regression over four ordinal compatibility classes (Poor,

   Neutral, Good, Excellent), trained on synthetically generated prole pairs:

   exp(wTs)

   E. Scholarship Finding Agent

   The Scholarship Agent enables international students to dis-cover relevant funding opportunities based on their academic

   L

   P (k | s) = k (7)

   j

   K

   j=1

   exp(wTs)

   prole and temporal constraints, querying a structured SQLite database of scholarships indexed by name, location, university,

   where s = [s1,…, s6] and K=4. The nal compatibility score is:

   S(A, B) = wTs (8)

   Personality descriptions are encoded into 768-dimensional SBERT vectors, providing a semantic re-ranking signal:

   provider, grant type, eligibility criteria, deadline, and grant value.

   Prole-Aware Query Construction. The agent exposes a RESTful endpoint accepting optional parameters for country, eld of study, grant type, and university. When parameters are absent, the authenticated users destination country and

   simsem(A, B) = vA · vB

   /lvA/l/lvB/l

   (9)

   program eld are substituted from their registration prole, ensuring personalized defaults without explicit input.

   Candidates scoring below =0.5 are ltered out, with the top-10 results returned in descending order. The content-based design avoids the cold-start problem inherent in collaborative ltering [13].

   D. Emotional Support Agent

   Relocation is inherently stressful. The Emotional Support Agent provides empathetic assistance to students navigating homesickness, culture shock, academic stress, and loneliness. This agent is built on affective computing principles that focuses on delivering sentimental responses [17].

   Intent Detection. Emotional intent is identied by the central LangChain Router via zero-shot intent classication [1], which infers affective intent from query semantics without relying on xed keyword lists. This ensures generalization to paraphrased or implicitly emotional expressions that rigid lexical triggers would miss.

   Sentiment Classiction. Queries routed to this agent are passed to a ne-tuned RoBERTa classier [12], [18] that maps the input into one of four affective states: neutral, mild distress, moderate distress, severe distress. The classication condence score pe is computed as:

   pe = softmax(We · hCLS + be) (10) where hCLS is the RoBERTa [CLS] token embedding and

   We, be are learned classication parameters.

   Context-Aware Response Generation. A rolling window of the ten most recent conversation turns is injected into the prompt as alternating User/Assistant exchanges. When pe exceeds a distress threshold e, the Routing Agent appends a compassion directive to the system prompt, ensuring responses are both factually grounded and emotionally appropriate [17], [18].

   Output. The agent returns a structured triple:

   Oe = (r, idagent, R) (11)

   where r is the generated response, idagent is the agent iden-tier, and R denotes a set of surfaced resource categories (mental-health-resources, emotional-support-guides), enabling the frontend to present professional contact information con-textually.

   Semantic Retrieval via TF-IDF and NER. Rather than relying on rigid pattern matching, the agent employs Named Entity Recognition (NER) to extract structured entities eld of study, institution type, funding category from the users academic prole. These entities are matched against the scholarship corpus using TF-IDF weighted retrieval:

   ( i

   TF-IDF(t, d, D) = tf(t, d) · log |D| (12)

   |{d D : t d}|

   where t is an NER-extracted term from the user prole, d is a scholarship document, and D is the full corpus. This ensures retrieval is driven by semantic relevance rather than exact string overlap.

   Temporal Filtering. A two-stage deadline lter retains only scholarships whose deadlines satisfy:

   tnow tstart + 365 (13)

   where tnow is the current date and tstart is the users program start date, ensuring all returned opportunities are temporally actionable. Grant values undergo ISO 4217 currency parsing to extract numeric amounts for consistent downstream com-parison.

   Output. Results are mapped to a structured schema contain-ing title, country, eld of study, deadline, amount, currency, eligibility type, grant type, university, provider, and an ex-ternal link surfaced directly within the chat interface for immediate access..

   F. Location and Event Recommender

   The Location Agent integrates spatial APIs with the CAPRI context-aware recommendation framework [14] and neural collaborative ltering [15]. Rather than returning random sug-gestions, the agent clusters university campus coordinates and student venues using DBSCAN [16]. A data point p qualies as a core cluster point when:

   |NE(p)| MinPts (14)

   The agent then recommends safe neighborhoods and culturally relevant events within these clusters based on the inferred user preference prole. This agent is in the active development phase.

5. Implementation and Results

   Abroad Compass is developed as a fully workable cross-platform application and evaluated by the system demon-stration on all ve specialized agents. The main interface components are presented in the following subsections.

   1. Authentication and User Onboarding

      At sign-up, the authentication layer collects detailed prole metadata including destination country, eld of study, program start date, budget range, and housing preferences. This prole is then used by the Scholarship and Roommate agents as default query context which eliminates redundant user input across sessions.

      Fig. 2. Login page with password encryption.

      Fig. 3. User registration form for capturing user prole attributes for personalized agent responses across the platform.

   2. Central Chat Interface and Agent Routing

      All user interaction happens through a single chat interface. The LangChain Router transparently directs queries to the correct specialized agent, and the active agent identier is surfaced in the top-right of the interface (e.g., Flight Finder, Legal Help, Scholarships). The sidebar navigation allows direct access to each agent module and displays the destination context of the authenticated user (destination country and eld of study).

   3. Legal Guidance Agent

      Figure 6 shows the Legal RAG Agent answering a query about documentation for a German study visa. The an-swer is based on chunks retrieved from the ofcial im-migration PDFs (gerenal_info_germany_visa.pdf,

      Fig. 4. Abroad Compass Chat interface, showing the system welcome message and enumerating all available agent capabilities.

      Fig. 5. Personalized agent modules and user context displayed in the navigation drawer (destination and eld of study).

      infostudents-data-germany-india.pdf) with the source citations directly below the response.

      Fig. 6. Legal RAG Agent answering a query about German study visa documentation, citing retrieved source PDFs inline.

   4. Flight Fare Forecasting Agent

      Figure 7 shows the Flight Fare Agent making a price prediction for the DEL FRA route. The correct cities are selected based on the user preferences lled in the sign up form, if no user specications are given. The response contains a predicted price range ($591-$722), a condence score (70%), a recommended booking window (2-4 weeks in advance), and live offers retrieved via the Amadeus API cross-validated against the model prediction. The transition from an emotional support query to a ight query in the same session demonstrates the routers ability to switch agent context dynamically without losing conversational continuity.

      Fig. 7. Flight Fare Agent delivering a DELFRA price prediction with Amadeus live offer reconciliation. The previous turn depicts dynamic agent switching from Emotional Support in the same session.

   5. Emotional Support Agent

      As shown in Figure 8, the Emotional Support Agent responds in a caring way to the query, I feel stressed. It correctly identies relocation-related anxiety based on the earlier conversation about legal issues. The re-sponse acknowledges the users feelings, normalizes the experience, and encourages further discussion. This ap-proach aligns with the compassion guideline added by the router when it detects emotional intent. Resource categories, such as mental-health-resources and emotional-support-guides, are included in the re-sponse metadata.

      Fig. 8. Emotional Support Agents response to a query with contextually grounded empathetic output, following a legal guidance exchange in the same session.

   6. Scholarship Finding Agent

      Figure 9 shows the Scholarship Recommender ltering out suitable scholarship sources for an Engineering student traveling to Germany, employing defaults based on user prole information without specifying any explicit lter settings. Figure 11 illustrates the specialized scholarship browser that supports multiple attribute lters, including country, eld, type of grant and university, resulting in output such as the Deutschlandstipendium

   7. Roommate Matching Agent

   Figures 12 and 13 illustrate the Roommate Matching Agent workow. Users ll a structured lifestyle prole that includes budget range, housing type, cleanliness level, smoking and pet

   Fig. 9. Scholarship Recommender agent identifying prole-matched funding opportunities via the chat interface, with a direct link to the scholarship browser.

   Fig. 10. Scholarship browser displaying ltered results for Engineering stu-dents in Germany, including grant value, deadline, institution, and eligibility type.

   preferences, study habits, sleep schedule, personality traits, and interests. The SBERT-augmented matching pipeline re-turns ranked candidates with weighted compatibility scores; as pr the given tests, the top match achieved a 97% compatibility score across ve shared interest categories, with subsequent matches at 82% and 80%. This reects meaningful differenti-ation while scoring the matches.

6. Conclusion and Future Work

Abroad Compass provides an altogether different agent infrastructure in comparison to monolithic chatbots. Through the utilization of RAG for legal correctness [4], [5], time-series regression for fare estimation [7], semantic matching for social compatibility [11], and affective computing for emotional intelligence [17], we provide a holistic relocation assistant experience. The cloud native architecture ensures that there is scalability and high availability.

Modularization ensures that specic agents can be upgraded or swapped out without affecting the core router functionality. Some potential improvements that could be done include:

(1) Increasing the volume of documents related to other destinations within the legal document corpus; (2) Automate the intake process for scholarship and immigration policies through live updates; (3) Utilizing multiple modalities for input within the agent pipeline; (4) Complete implementation of location and event nder agent; and (5) Utilizing cloud infrastructure to increase availability and dynamic scaling as user trafc increases.

Fig. 11. (Extended) Scholarship browser displaying ltered results for Engineering students in Germany, including grant value, deadline, institution, and eligibility type.

Fig. 12. Roommate prole form capturing lifestyle attributes including cleanliness level, study habits, sleep schedule, personality traits, and interests used for SBERT-augmented compatibility scoring.

References

1. S. Yao et al., ReAct: Synergizing Reasoning and Acting in Language Models, arXiv preprint arXiv:2210.03629, 2022.

2. Q. Wu et al., AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation, arXiv preprint arXiv:2308.08155, 2023.

3. J. S. Park et al., Generative Agents: Interactive Simulacra of Human Behavior, in Proc. ACM UIST, 2023.

4. I. Chalkidis et al., LEGAL-BERT: The Muppets Straight Out of Law School, in Findings of EMNLP, 2020.

5. V. Karpukhin et al., Dense Passage Retrieval for Open-Domain Ques-tion Answering, in Proc. EMNLP, 2020.

6. P. Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, in Advances in Neural Information Processing Systems (NeurIPS), vol. 33, pp. 94599474, 2020.

7. H. Shu et al., Deep-Learning-Powered GRU Model for Flight Ticket Fare Forecasting, Applied Sciences, vol. 13, no. 10, p. 6032, 2023.

8. X. Sun et al., Civil Airline Fare Prediction with a Multi-Attribute Dual-Stage Attention Mechanism, Applied Intelligence, 2021.

9. K. Tziridis et al., Airfare Prices Prediction Using Machine Learning Techniques, in Proc. IEEE EAIS, 2020.

10. S. Hochreiter and J. Schmidhuber, Long Short-Term Memory, Neural Computation, vol. 9, no. 8, pp. 17351780, 1997.

11. N. Reimers and I. Gurevych, Sentence-BERT: Sentence Embeddings Using Siamese BERT-Networks, in Proc. EMNLP, 2019.

12. Y. Liu et al., RoBERTa: A Robustly Optimized BERT Pretraining Approach, arXiv preprint arXiv:1907.11692, 2019.

13. W. Fan et al., Graph Neural Networks for Social Recommendation, in

    Proc. WWW, 2019.

14. K. Yao et al., CAPRI: Context-Aware Interpretable Point-of-Interest Recommendation Framework, arXiv preprint arXiv:2306.11395, 2022.

15. X. He et al., Neural Collaborative Filtering, in Proc. WWW, 2017.

16. M. Ester, H.-P. Kriegel, J. Sander, and X. Xu, A Density-Based Algorithm for Discovering Clusters in Large Spatial Databases with Noise, in Proc. 2nd Int. Conf. on Knowledge Discovery and Data Mining (KDD-96), pp. 226231, 1996.

    Fig. 13. Roommate match results ranked by weighted compatibility score (97%, 82%, 80%), displaying shared interests, budget range, and housing type for each candidate.

17. H. Rashkin et al., Towards Empathetic Open-domain Conversation Models: A New Benchmark and Dataset, in Proc. ACL, 2019.

18. W. Shen et al., TED: Turn Emphasis with Dialogue Feature At-tention for Emotion Recognition in Conversation, arXiv preprint arXiv:2501.01123, 2020.

19. P. Zhang, G. Zeng, T. Wang, and W. Lu, TinyLlama: An Open-Source Small Language Model, arXiv preprint arXiv:2401.02385, 2024.

20. J. Sweller, Cognitive Load During Problem Solving: Effects on Learn-ing, Cognitive Science, vol. 12, no. 2, pp. 257285, 1988.

21. International Organization for Migration, World Migration Re-port 2024, IOM, Geneva, 2024. [Online]. Available: https:// worldmigrationreport.iom.int

22. M. Jones, J. Bradley, and N. Sakimura, JSON Web Token (JWT), IETF RFC 7519, May 2015. [Online]. Available: https://tools.ietf.org/ html/rfc7519

23. G. Molna´r and Z. Szu¨ts, The Role of Chatbots in Formal Education, in Proc. IEEE SACI, 2018.

24. P. Smutny and P. Schreiberova, Chatbots for Learning: A Review of Educational Chatbots for the Facebook Messenger, Computers & Education, vol. 151, 2020.

25. R. Santos et al., LAWBOTS: Utilization of AI Chatbots for Legal Advising in the Philippines, in Proc. IEEE, 2024.

26. N. Haidari et al., A Proposed Cognitive Framework Model for a Student Support Chatbot in a Higher Education Institution, Academia, 2022.

27. A. Dixit et al., Personality-Aware Roommate Recommendation Sys-tem, 2018.