EcoSort AI: A C2B Mobile Ecosystem for AI-Driven E-Waste Management

EcoSort AI: A C2B Mobile Ecosystem for AI-Driven E-Waste Management

Tushar Kumar

Dept. of Computing Technologies, School of Computing SRM Institute of Science & Technology Kattankulathham, Chennai,

India tushaarxr@gmail.com

Brindha R

Dept. of Computing Technologies, School of Computing SRM Institute of Science & Technology Kattankulathham, Chennai, India brindhar@srmist.edu.in

N. A. S. Vinoth

Dept. of Computing Technologies, School of Computing SRM Institute of Science & Technology Kattankulathham, Chennai, India. vinoths@srmist.edu.in

Tanishq Agrawal

Dept. of Computing Technologies, School of Computing SRM Institute of Science & Technology Kattankulathham, Chennai, India mahamiatanishq@gmail.com

K. M. Umamaheswari

Dept. of Computing Technologies, School of Computing SRM Institute of Science & Technology Kattankulathham, Chennai, India umamahek@srmist.edu.in

R. K. Pongiannan

Dept. of Computing Technologies, School of Computing SRM Institute of Science & Technology Kattankulathham, Chennai, India. pongiank@srmist.edu.in

AbstractEcoSort AI is a cross-platform mobile application (iOS, Android, Web) that addresses the global electronic waste (e-waste) crisis by integrating AI-powered image recognition, a Consumer-to-Business (C2B) marketplace, and a corporate exchange programme aggregator into a single, unified platform. The AI engine employs Google Gemini 2.5 Flash with a structured JSON prompt chain to classify e-waste items, infer material composition, assign a hazard level (Low / Medium / High), flag data-security risk, and generate device-specific disposal guidance

• all from a single photograph. Beyond identification, EcoSort AI introduces a dual-role marketplace in which consumers create multi-image listings and admin-verified vendors express interest through a lead-generation pipeline enforced by PostgreSQL Row Level Security (RLS). An Exchange Programs Hub aggregates trade-in offers from major brands (Apple, Samsung, Amazon, Flipkart) within a single browsable portal. An offline-first cache manager with a persistent sync queue ensures functionality in low-connectivity environments. The platform comprises 35 screens across three route groups, a companion web-based admin panel for vendor verification and listing moderation, and a comprehensive input-validation and rate-limiting utility layer. Built on React Native 0.81.5 / Expo SDK 54, Clerk JWT authentication, and Supabase (PostgreSQL) with per-table RLS policies, EcoSort AI delivers a production-grade architecture aligned with United Nations Sustainable Development Goal 12 (Responsible Consumption and Production).

  Index TermsE-waste management, mobile application, Gem- ini 2.5 Flash, C2B marketplace, circular economy, Supabase RLS, LMM prompt engineering, offline-first, SDG 12, React Native

  1. INTRODUCTION

     The exponential rise in consumer electronics has led to a parallel surge in electronic waste (e-waste), now recognised

     as the fastest-growing domestic waste stream globally. The Global E-Waste Monitor 2020 reported over 53.6 million met- ric tonnes generated in 2019, a figure projected to surpass 74 million tonnes by 2030 [1]. Improper disposal of such waste, laden with toxic materials such as mercury, cadmium, and lead, poses severe environmental and public-health hazards [11]. Despite growing awareness, individuals frequently face four distinct knowledge gaps: (i) they do not know what hazardous materials their devices contain; (ii) they lack guidance on safe disposal; (iii) they cannot locate certified recycling facilities; and (iv) they are unaware that their old devices may carry residual monetary value.

     EcoSort AI is designed as a closed-loop platform that ad- dresses all four gaps within a single mobile application. The AI engine, built on Google Gemini 2.5 Flash [2] with a structured multi-field JSON prompt chain, analyses a photograph of any electronic item and returns a complete ScanResult including item type, material composition, hazard level, data- security risk flag, recycling value assessment, and step-by-step disposal instructions. The C2B Marketplace then converts this awareness into action by connecting consumers with admin- verified vendors, while the Exchange Programs Hub surfaces corporate trade-in offers from Apple, Samsung, Amazon, and Flipkart in a single aggregated portal. A GPS-driven recycling centre locator using the Haversine formula closes the loop by directing users to certified drop-off points.

     The platform comprises 35 screens across three route groups (authentication, consumer, vendor), a companion web-based admin panel for vendor verification and listing moderation,

     and an offline-first architecture with a persistent sync queue making the application viable in low-connectivity environ- ments typical of developing-country e-waste scenarios.

  2. Novel Contributions

     Prior work on e-waste mobile platforms reveals a consistent segmentation: identification tools lack marketplaces, market- place tools lack AI, and none integrate offline resilience, dual- role access control, and quantified environmental tracking into a single application. EcoSort AI makes the following distinct contributions:

     1. Structured JSON Prompt Chain for LMM- Based E-Waste Classification. We design and implement a domain-specific two-schema prompt architecture (Algorithm 1) that constrains Google Gemini 2.5 Flash via responseMimeType: "application/json" and temperature 0.6 to produce structured ScanResult objects. The schema differentiates e-waste items (returning eight typed fields including hazardLevel, dataSecurityRisk, and recyclingValue) from non-e-waste items (returning a funMessage). A fallback retry mechanism handles non-compliant outputs. This design eliminates the need for labelled training datasets and domain-specific CNN retraining while generalising across the full variability of consumer electronics.

     2. Trust-Tiered C2B Marketplace with RLS- Enforced Data Isolation. We architect a two- sided marketplace comprising a 1,331-line service layer (marketplaceService.ts) and five interconnected database tables (listings, vendors, vendor_interests, transactions, exchange_offers). Vendor onboarding requires admin verification via a companion web panel; the is_verified flag, set exclusively by an admin role, gates all vendor-exclusive database operations. Per-table RLS policies enforce strict data isolation: consumers may only mutate their own records, and only verified vendors may INSERT into vendor_interests.

     3. Offline-First Architecture with Persistent Sync Queue. We implement a singleton OfflineCache manager (529 lines) using AsyncStorage as the persistence backend, with a NetworkManager observer that automatically drains a persisted SyncQueueItem[] on reconnection. Cache capacity limits (100 scan results, 500 recycling centres, 200 disposal guides) and a 24-hour expiry prevent stale data without requiring permanent connectivity a design choice with direct relevance to e-waste management in semi-urban and rural contexts.

     4. Exchange Programs Hub. We design and implement a centralised aggregation layer (exchangeService.ts, 359 lines) that surfaces corporate trade-in offers with automatic view_count and click_count analytics. Contextual offer retrieval (getOffersForItem()) matches a scanned items

        category and brand to relevant trade-in programmes, creating a direct pipeline rom AI identification to monetised disposal.

     5. Comprehensive Input Validation and Rate- Limiting Layer. A standalone validation module (validation.ts, 292 lines) provides 12 typed validators covering email, password, image URI, Base64 encoding, GPS coordinates, scan result schema, and an in-memory rate limiter a security contribution absent from all comparable e-waste applications identified in the literature.

  3. LITERATURE REVIEW AND COMPARATIVE

     ANALYSIS

     1. The Growing Threat of E-Waste

        The Global E-Waste Monitor 2020 estimated over 53.6 million metric tonnes of e-waste generated in 2019, projected to surpass 74 million tonnes by 2030 [1]. E-waste contains valuable resources such as gold and copper alongside haz- ardous materials including mercury and lead, making its mis- management simultaneously an economic loss and a public- health crisis [11].

     2. Current E-Waste Management Practices

        Conventional e-waste management relies on manual collec- tion, community drives, or municipal drop-off points. While Extended Producer Responsibility (EPR) legislation has been introduced globally, its on-ground execution remains incon- sistent due to limited user awareness and the absence of consumer-facing financial incentives [10]. Existing platforms rarely connect consumers directly with verified recyclers through a transparent, application-level marketplace.

     3. Role of AI in Waste Classification

        Deep learning has been applied to classify municipal waste using Convolutional Neural Networks, notably WasteNet [7] and SmartBin [8]. However, CNN-based approaches require large, domain-specific labelled datasets which remain scarce for the heterogeneous e-waste category. Large Multimodal Models (LMMs) such as Google Gemini 2.5 Flash [2] gener- alise across device types, infer component-level material com- position from visual cues (port types, form factor, branding), and produce natural-language disposal guidance capabili- ties beyond discriminative classifiers.

     4. Mobile Applications for Sustainability

        Platforms such as ecoATM [9], RecycleNation, and iRecycle offer recycler directories or kiosk-based device buyback. Table I presents a structured feature comparison of EcoSort AI against these platforms and WasteNet [7]. EcoSort AI is the only platform to simultaneously provide LMM-based AI identification, hazard-level output, a two-sided verified C2B marketplace, an exchange programme aggregator, environmen- tal impact scoring, an offline mode, and a companion admin panel.

        TABLE I

        Feature Comparison of E-Waste Platforms

        Feature
        LMM/AI image classification			×	×
        Hazard level & data-risk flag		×	×	×
        C2B marketplace (two-sided)		×	×	×
        Admin-verified vendor flow		×	×	×
        Exchange programme aggregator		×	×	×
        GPS recycler map locator
        Environmental impact score		×	×	×
        Offline-first with sync queue		×	×	×
        Cross-platform (iOS/Android)		×
        Companion admin web panel		×	×	×
        Input validation & rate limit		×	×	×

     5. Research Gaps

     No existing platform unifies LMM-based e-waste identifi- cation, a trust-tiered C2B marketplace, exchange programme aggregation, offline-first resilience, and a companion admin panel in a single mobile application. EcoSort AI is architected to fill this gap end-to-end.

  4. SYSTEM ARCHITECTURE

     1. Technology Stack

        EcoSort AI is built on a modern, fully open-source- compatible stack detailed in Table II. The frontend uses React Native 0.81.5 with Expo SDK 54 and file-based routing via Expo Router 6. UI components are provided by React Native Paper 5.13 (Material Design 3), with animations by React Native Reanimated 4.1. Authentication is handled by Clerk v2.19 with JWT tokens cached in Expo SecureStore. The database layer uses Supabase (PostgreSQL) with per- table RLS policies enforced via JWT claim extraction. AI processing is performed by Google Gemini 2.5 Flash via the @google/generative-ai SDK v0.24.1.

        TABLE II

        Production Technology Stack EcoSort AI v1.0.0

        Component

        Technology (Version)

        Mobile framework Routing Language UI library Animation engine Authentication Database File storage AI engine AI SDK Maps & GPS Offline storage Network detection Admin panel

        React Native 0.81.5 / Expo SDK 54 Expo Router 6.x (file-based, typed) TypeScript 5.7+ React Native Paper 5.13 (MD3) React Native Reanimated 4.1 Clerk v2.19 (JWT + SecureStore) Supabase (PostgreSQL + RLS) Supabase Storage (S3-compatible) Google Gemini 2.5 Flash @google/generative-ai v0.24.1 expo-location v19.0.8 AsyncStorage 2.2.0 @react-native-community/netinfo v11.4.1 Vanilla HTML/CSS/JS + Supabase JS

     2. Three-Tier Architecture

        The system follows a three-tier client-server model with a clear separation of concerns. The mobile client (React Native) communicates with Supabase over HTTPS for all database and storage operations, with every request carrying a Clerk-issued JWT that is validated at the Supabase gateway before any query executes. Image payloads are forwarded directly from the client to the Gemini 2.5 Flash endpoint as base64-encoded JPEG data, and the structured JSON response is persisted in Supabase under the authenticated users RLS scope. GPS coordinates are resolved client-side via expo-location and the Haversine formula to rank nearby recycling centres without a dedicated geospatial backend service.

        The provider hierarchy in the root layout (_layout.tsx) is: ClerkProvider ThemeProvider SafeAreaProvider PaperProvider Slot, ensuring authentication context, theme, and Material Design

        components are available to all 35 screens.

     3. Navigation and Screen Architecture

        The application is structured across three Expo Router route groups and 35 total screens:

        • (auth) Stack navigator: Login, Register, Password Reset (unauthenticated access).

        • (main) Tab navigator with 5 visible tabs (Scan, Repository, Recycling Centers, Marketplace, Exchange) and 2 hidden routes (Profile, Search). Accessible to authenticated consumers.

        • (vendor) Stack navigator: Register, Pending, Dash- board, Browse, My Interests. Accessible to authenticated vendors after admin verification.

     4. Service Layer Architecture

     All business logic is encapsulated in src/api/, fol- lowing four consistent design patterns: (i) snake case camelase transformation on all DB responses; (ii) grace-

     ful degradation returning null/[] on errors rather than throwing; (iii) fire-and-forget for non-critical increments (view counts, interest counts); and (iv) ownership verification via

     .eq(user_id, userId) on all mutation queries. Ta- ble III summarises the service files.

     TABLE III Service Layer Summary

     File	Lines	Responsibility
     geminiService.ts	179	Gemini 2.5 Flash AI pipeline
     	375	Scans, centers, e-waste items
     supabaseService.ts
     marketplaceService.ts	1,331	Listings, vendors, interests
     exchangeService.ts	359	Exchange offers & analytics
     apiUtils.ts	76	Connectivity diagnostics

  5. Methodology

     1. LMM Prompt Engineering for E-Waste Classification

        Rather than training a bespoke CNN, EcoSort AI exploits the zero-shot generalisation capacity of Google Gemini 2.5 Flash via a carefully engineered dual-schema prompt chain.

        This design is deliberate for three reasons: (i) labelled e- waste image datasets sufficient for CNN training are not publicly available at the required scale and diversity; (ii) LMM inference inherits Geminis continual model updates without retraining costs; and (iii) LMMs can produce natural-language disposal guidance, data-security warnings, and recycling-value assessments in a single inference pass outputs that discrim- inative classifiers cannot generate.

        The system prompt instructs the model to act as EcoSort AI, a friendly e-waste expert. Structured output is enforced via the responseMimeType: "application/json" parameter and a temperature of 0.6. The API call uses a maximum of 2,048 output tokens. The prompt defines two conditional response schemas:

        1. E-waste schema (isEWaste: true): returns itemType, materials[], hazardLevel (low | medium | high), disposalMethod, recyclingValue, dataSecurityRisk (boolean), and confidence (low | medium

           | high).

        2. Non-e-waste schema (isEWaste: false): returns a

           funMessage a sarcastic but friendly comment (e.g.,

           Unless you plan to charge your phone with cake. . . ).

           Algorithm 1 formalises the complete inference pipeline.

           Algorithm 1 Gemini 2.5 Flash E-Waste Classification Pipeline

           Require: Image URI U , system prompt Psys

           Ensure: ScanResult object S

           1: B readAsStringAsync(U , base64)

           2: Strip data-URI prefix from B

           3: Construct inline image part Ipart {mimeType: im- age/jpeg, data: B}

           4: Set API params: temperature=0.6,

     2. Dual-Flow C2B Marketplace

        The marketplace is architected around two strictly separated roles enforced at both the application and database layers.

        1. Consumer Flow: Consumers create listings with ti- tle, category, brand, model, condition (working | mi- nordamage | notworking | forparts), asking price, negotiability flag, and up to multiple images uploaded to the

           listing-images Supabase Storage bucket. Listings auto- expire after 30 days (expires_at column). The consumer dashboard (my-listings.tsx) displays view counts, in- terest counts, and the status lifecycle: Draft Active Sold

           / Expired / Removed.

        2. Vendor Flow: Vendors register by submitting business details (name, type, phone, address, service areas, categories of interest, and verification documents). Documents are up- loaded to a private Supabase Storage bucket with 1-year signed URL expiry. The registration record is created with verification_status: "pending" and reviewed by an admin via the companion web panel. On approval, the admin sets status to "verified". Only verified ven- dors may INSERT into vendor_interests, enforced by an RLS policy. A UNIQUE constraint on (listing_id, vendor_id) prevents duplicate interest submissions. Upon a consumer accepting an interest, contact_revealed is set to true, exposing the consumers contact details to the vendor and creating a transactions record.

     3. Exchange Programs Hub

        The exchangeService.ts module provides contex- tual offer retrieval via getOffersForItem(), which matches a ScanResults category and brand against the exchange_offers tables categories (JSONB) and brands_accepted (JSONB) columns. Each offer car-

        ries min_value, max_value, valueText (e.g., Up to

        maxTokens=2048, responseMimeType=application/json

        25,000), offerUrl, and isFeatured flag. Click-through

        5: Rraw GeminiFlash.generateContent(Psys,

        Ipart)

        6: if isValidJSON(Rraw) then 7: R parseJSON(Rraw) 8: else

        9: Rraw GeminiFlash.generateContent(Pretry,

        Ipart)

        10: R parseJSON(Rraw); fallbackParsed

        true

        11: end if

        12: Validate hazardLevel {low, medium, high} 13: Validate confidence {low, medium, high} 14: S buildScanResult(R, timestamp, userId)

        15: supabase.insert(scans, S) {RLS-scoped to user}

        16: if offline then

        17: OfflineCache.enqueueSyncItem(S)

        18: end if

        19: return S

        events are tracked via trackOfferClick().

     4. Recycling Centre Locator

        Device GPS is obtained via expo-location. The getNearbyRecyclingCenters() function retrieves up to 50 active, verified centres from Supabase and ranks them by Haversine distance. Each centre record includes accepts_data_devices, data_wiping_service, and pickup_service boolean flags, providing nuanced guidance to users with privacy-sensitive devices.

     5. Offline-First Architecture

        The OfflineCache singleton (529 lines) implements three complementary patterns. The Singleton Pattern en- sures a single cache manager across the application life- cycle. The Observer Pattern allows NetworkManager to notify listeners on connectivity state changes, triggering au- tomatic SyncQueue drainage. The Queue Pattern persists SyncQueueItem[] to AsyncStorage with a retry policy of 3 attempts at 2-second intervals. Cache entries expire after 24 hours. Table IV details capacity limits per cache type.

        TABLE IV

        Offline Cache Capacity Configuration

        Cache Type	Max Items	Expiry
        Scan results	100	24 h
        Recycling centres	500	24 h
        Disposal guidance	200	24 h

     6. Security and Privacy Architecture

        EcoSort AI employs a four-layer security model:

        1. JWT Validation. Every Supabase request carries a Clerk-issued JWT injected via a custom fetch interceptor in useSupabase.ts. The Supabase gateway validates the token signature and extracts the sub claim via:

           coalesce(

  6. DATABASE DESIGN

     The Supabase PostgreSQL schema comprises eight tables across two SQL migration files (supabase_setup.sql and supabase_marketplace.sql). Core tables include scans (storing AI analysis results per user with JSONB materials arrays), recycling_centers (with lat- itude/longitude decimals, JSONB accepts_items and certifications arrays, and boolean service flags), and user_preferences (notification, location, andunit pref- erences per user).

     The marketplace schema introduces listings (with JSONB images array, 30-day auto-expiry, and fire-and-forget view_count/interest_count fields), vendors (with JSONB service_areas, categories_interested, and documents arrays and a 2-decimal rating field),

     current_setting(request.jwt.claims,true)vendor_interests (with a UNIQUE(listing_id,

     ::json->>sub,

     auth.jwt()->>sub

     )

     1. Row Level Security. All eight database tables have RLS enabled. Table V summarises the per-table policy matrix.

        TABLE V

        Table	Operation	Policy
        scans recycling_centers user_preferences listings listings vendors vendors vendor_interests vendor_interests transactions exchange_offers	All CRUD SELECT All CRUD SELECT INSERT/UPDATE/ DELETE SELECT INSERT/UPDATE SELECT INSERT SELECT SELECT	Own records only Anyone (active) Own records only Anyone (active) + own Own records only Verified + own Own records only Listing owner + vendor Verified vendors only Seller or vendor party Anyone (active only)

        Row Level Security Policy Matrix

     2. Admin-Controlled Vendor Verification. The verification_status field is mutated exclusively by the admin web panel using an admin-role Supabase client, preventing self-elevation attacks.

     3. Input Validation and Rate Limiting. The validation.ts module (292 lines) provides 12 typed validators including Base64 image validation (up to 10 MB), GPS coordinate range checking (±90/±180), scan result schema validation, and an in-memory rate limiter with configurable window and threshold parameters.

     1. Admin Dashboard

     A companion web panel (Vanilla HTML/CSS/JS with the Supabase JS client) provides four management pages: Login, Dashboard (platform statistics), Users (profile management, vendor approval/rejection), and Listings (moderation). This panel is the exclusive interface for vendor verification, a deliberate architectural decision that removes self-registration as an attack vector.

     vendor_id) constraint and an ai_moderation_score column for future AI-assisted message moderation), transactions (tracking payment method, status, and mutual seller/vendor ratings), and exchange_offers (with JSONB categories and brands_accepted arrays for contextual matching).

  7. RESULTS AND EVALUATION

     1. AI Classification Evaluation

        To evaluate the Gemini 2.5 Flash prompt-chain pipeline, we constructed a benchmark of 50 consumer electronics images spanning seven device categories: smartphones, laptops, bat- teries, printers, monitors, cables/adapters, and miscellaneous peripherals. Images were sourced from publicly available e- waste image repositories and supplemented with field-captured photographs during functional testing. Each image was pro- cessed through Algorithm 1 and evaluated against ground- truth labels assigned by two independent annotators (inter- rater agreement Cohens = 0.91, indicating near-perfect agreement).

        Table VI reports per-category precision, recall, and F1 scores. The system achieved a macro-averaged F1 of 0.87. Smartphones achieved the highest F1 (0.94) due to their distinctive and consistent visual form factor. Miscellaneous peripherals scored lowest (0.76) owing to their high visual heterogeneity consistent with the known limitation of zero- shot LMM inference on ambiguous inputs. Hazard-level clas- sification (Low / Medium / High) achieved an overall accuracy of 89%, with primary confusion occurring between Medium and High for battery-containing devices, where the presence of a lithium cell is not always visually distinguishable. The fallbackParsed flag was triggered in 4% of inference calls, confirming that the retry mechanism successfully recov- ers from non-compliant JSON outputs.

     2. Usability Evaluation

        A System Usability Scale (SUS) evaluation [12] was con- ducted with 20 participants (12 male, 8 female; ages 18 35; mix of technology-proficient and general users) recruited from the university campus and surrounding community.

        TABLE VI

        AI Classification Performance by Device Category (n = 50)

        Category	Precision	Recall	F1
        Smartphones	0.95	0.93	0.94
        Laptops	0.91	0.90	0.91
        Batteries	0.88	0.86	0.87
        Printers	0.86	0.85	0.86
        Monitors	0.89	0.87	0.88
        Cables / Adapters	0.82	0.80	0.81
        Misc. Peripherals	0.78	0.74	0.76
        Macro Avg.	0.87	0.85	0.87

        Participants completed five standardised tasks: (T1) account registration via Clerk; (T2) scanning an e-waste item with the Gemini pipeline; (T3) creating a marketplace listing with image upload; (T4) locating a recycling centre on the GPS map; and (T5) browsing the Exchange Programs Hub. Task completion times and binary success outcomes were recorded, followed by the 10-item SUS questionnaire.

        Table VII summarises results. The mean SUS score of 82.4 corresponds to a Grade-B (Good) usability rating on the Bangor et al. adjective scale. Task T2 (AI scan) recorded the lowest completion rate (85%) due to occasional retry prompts when image quality was insufficient for the Gemini API. This finding directly informed a planned UX enhancement: real- time image quality feedback before scan submission, listed in the future roadmap.

        TABLE VII

        SUS Usability Study Results (n = 20 Participants)

        Task	Completion	Mean Time (s)
        T1: Account registration (Clerk)	100%	48
        T2: Scan e-waste item (Gemini)	85%	34
        T3: Create marketplace listing	90%	62
        T4: Locate recycling centre (GPS)	95%	22
        T5: Browse Exchange Programs Hub	100%	18
        Mean SUS Score		82.4 (Grade B)

     3. System Performance and Scalability

        Supabases PostgreSQL backend supports horizontal read scaling via PgBouncer connection pooling. Image uploads are routed directly to Supabase Storage (S3-compatible object store), decoupling storage I/O from query throughput. The Gemini 2.5 Flash API is stateless and horizontally scalable; rate limits (60 req/min on free tier, 1,000+ req/min on pro- duction tiers) are managed client-side with exponential back- off implemented in apiUtils.ts. Table VIII reports key latency figures observed during functional load testing.

        TABLE VIII

        System Performance Benchmarks

        Operation	Mean	95th Pct.
        Gemini 2.5 Flash round-trip Marketplace feed (20 listings) Image upload to Supabase ( 3 MB) Map load (50 recycling centres) Clerk JWT refresh Offline cache read (AsyncStorage)	2.1 s 0.38 s 1.1 s 0.55 s 0.12 s 0.04 s	3.8 s 0.71 s 2.0 s 0.94 s 0.21 s 0.09 s

        The offline cache sustains full read functionality (scan his- tory, recycling centre list, disposal guides) with zero network latency. The sync queue successfully drained all pending write operations within 2.3 seconds (mean) of network reconnection in connectivity-toggle tests.

     4. Environmental Impact Analysis

     Each confirmed disposal event in EcoSort AI is instru- mented with a per-device CO2 diversion factor derived from published life-cycle assessment (LCA) literature [13], repre- senting the greenhouse-gas emission avoided when a device is

     recycled via a certified channel rather than landfilled or infor- mally processed. Table IX presents the per-category diversion estimates. These values are surfaced on each users personal dashboard, translating individual disposal actions into tangible ecological metrics.

     TABLE IX

     ESTIMATED CO2 DIVERSION PER DEVICE CATEGORY (LCA-DERIVED)

     Device Category	CO2-eq Diverted (kg)
     Smartphone	2.5
     Laptop / Tablet	5.8
     Battery (Li-ion)	1.2
     Monitor / TV	8.1
     Printer	4.4
     Cable / Adapter	0.4
     Misc. Peripheral	1.9
     Weighted Mean	3.2

  8. DISCUSSION

     The empirical results confirm that EcoSort AI achieves its design goals across all evaluated dimensions. The macro- averaged F1 of 0.87 (Section VII-A) demonstrates that a structured LMM prompt chain is a viable alternative to CNN retraining for e-waste classification, with the additional ad- vantage of producing natural- language disposal guidance, data-security warnings, and recycling-value assessments in a single inference pass. The 4% fallback-parsing rate confirms robustness of the retry mechanism.

     The SUS score of 82.4 (Section VII-B) validates the mobile- first, Material Design 3 interface philosophy. The lowest task completion rate (T2, 85%) is attributable to image quality constraints at the Gemini API boundary an infrastructure limitation, not a usability design failure and has been captured in the development roadmap.

     The RLS policy matrix (Table V) and admin-controlled vendor verification collectively ensure marketplace integrity without compromising the consumer experience. No cross- role data leakage occurred during the testing period. The offline sync queue demonstrated reliable operation in the connectivity-toggle scenario, confirming that the application is deployable in semi-urban and rural contexts where e-waste generation is significant but connectivity is intermittent.

     Limitations of the current evaluation include the modest benchmark size (n = 50), which was constrained by the scarcity of publicly labelled e-waste image datasets, and the

     single-institution composition of the usability cohort. Expand- ing both the benchmark and the participant pool to include diverse geographies and demographics is the primary target for v2.0 evaluation.

  9. CONCLUSION AND FUTURE WORK

EcoSort AI presents a production-grade, empirically evalu- ated mobile platform that addresses the e-waste crisis through five integrated mechanisms: LMM-based identification via Google Gemini 2.5 Flash, a trust-tiered C2B marketplace with RLS-enforced data isolation, an Exchange Programs Hub aggregating major brand trade-in offers, a GPS-driven recycling centre locator, and an offline-first architecture with persistent sync queue. The platforms 35-screen interface, companion admin panel, and comprehensive input-validation layer represent a complete software engineering contribution, not merely a prototype integration of third-party services. Alignment with UN SDG 12 is operationalised through per-

user CO2 diversion dashboards grounded in LCA literature.

Planned future enhancements include: (i) AI-driven price estimation using regression models trained on trade-in market data; (ii) in-app real-time messaging between vendors and consumers via Supabase Realtime WebSockets; (iii) carbon- footprint gamification with badges and leaderboards; (iv) doorstep e-waste pickup scheduling via logistics courier APIs (Shiprocket, Dunzo); (v) expansion of the classification bench- mark to 500+ images across 15 device categories; and (vi) multilingual interface support targeting regional Indian lan- guages.

1. Y. Zhao, X. Wang, H. Liu, and M. Yang, WasteNet: A Convolutional Neural Network for Waste Classification, in Proc. Int. Conf. Intelligent Computing, 2019, pp. 136145.

2. A. Kumar and A. Sharma, SmartBin: A Deep Learning Approach for Waste Detection and Classification, Int. J. Computer Applications, vol. 183, no. 30, pp. 2024, 2021.

3. ecoATM, ecoATM Electronics Recycling Kiosks, 2022. [Online].

   Available: https://www.ecoatm.com

4. Toxics Link, E-waste Management in India: Challenges and Opportu-

   nities, 2021. [Online]. Available: https://toxicslink.org

5. B. H. Robinson, E-waste: An Assessment of Global Production and Environmental Impacts, Science of the Total Environment, vol. 408, no. 2, pp. 183191, 2009.

6. A. Bangor, P. Kortum, and J. Miller, Determining What Individual SUS Scores Mean: Adding an Adjective Rating Scale, J. Usability Studies, vol. 4, no. 3, pp. 114123, 2009.

7. European Environment Agency, Life Cycle Assessment of ICT Equip- ment: CO2 Emission Factors for Consumer Electronics, EEA Technical Report, 2021. [Online]. Available: https://www.eea.europa.eu

ACKNOWLEDGMENT

The authors express sincere gratitude to Dr. C. Muthamizhchelvan, Vice-Chancellor, SRM Institute of Science and Technology, for the facilities provided and ongoing support; Dr. Leenus Jesu Martin M, Dean-CET; Dr. Revathi Venkataraman, Professor and Chairperson, School of Computing; Dr. M. Pushpalatha and Dr. Lakshmi, Associate Chairpersons, School of Computing; Dr. Niranjana G, Professor and Head, Department of Computing Technologies; faculty advisor Dr. Madhumitha K; and project guide Dr. Brindha R, Department of Computing Technologies, SRM Institute of Science and Technology, Kattankulathham, for her guidance, autonomy, and inspiration throughout this work.

REFERENCES

1. V. Forti, C. P. Balde´, R. Kuehr, and G. Bel, The Global E-waste Monitor 2020: Quantities, Flows, and the Circular Economy Potential, United Nations University / ITU / ISWA, 2020.

2. Google DeepMind, Gemini 2.5 Flash: Our Most Intelligent Gemini API Model, 2025. [Online]. Available: https://ai.google.dev/gemini-api/ docs/models/gemini

3. Supabase, Supabase Documentation: The Open Source Firebase Alter- native, 2024. [Online]. Available: https://supabase.com/docs

4. Clerk, Clerk Authentication Documentation, 2024. [Online]. Available:

   https://clerk.com/docs

5. United Nations, Sustainable Development Goal 12: Ensure Sustainable Consumption and Production Patterns, 2015. [Online]. Available: https:

   //sdgs.un.org/goals/goal12

6. Meta Platforms Inc., React Native Documentation, 0.81.5 ed., 2024. [Online]. Available: https://reactnative.dev