Projects
A lot of ideas, but some are still under construction!
Showing 13 of 13 projects
Task Turkey — Crowdsourced AI Video Platform
A micro-task platform where users contribute free-tier AI tool usage (Kling, Runway, Hailuo, Wan 2.1) to collectively produce monetizable YouTube videos. Users claim scene tasks, generate clips with AI tools, upload to their own Google Drive, and earn tokens. Zero storage cost — the platform stores only URL strings.
Key Learnings:
- Built a full-stack SaaS platform with atomic DB transactions and race-condition-safe task claiming
- Implemented Google OAuth + Drive API integration with open-redirect protection
- Designed a zero-cost architecture: token ledger, RLS policies, and URL-only video storage
Earn Rupai - Task Earning Platform
A production-ready platform with 150+ active students where users complete tasks (surveys, app installs, quizzes) and earn real money via UPI. Full stack built end-to-end: Flutter UI, Node.js backend, Supabase database, with wallet, referrals, and an admin panel. Published on Google Play Store.
Key Learnings:
- Built a complete product end-to-end with real users, not just a UI prototype
- Designed wallet ledger, referral, and task-completion systems with RLS-secured Supabase schema
- Handled production challenges: race conditions, payments, and user support
Trinetra - Mobile Application
A feature-rich mobile application available on Google Play Store. Designed to provide seamless user experience with modern UI/UX principles and robust functionality.
Key Learnings:
- Developed cross-platform mobile app with Flutter
- Implemented Material Design principles
- Managed app deployment and updates on Play Store
CodeSnap - Code Snippet Manager
Save, organize, and share code snippets with syntax highlighting. Features collections, tags, and search functionality for easy snippet management.
Key Learnings:
- Planning CRUD operations with RESTful API
- Designing search functionality with MongoDB
- Implementing syntax highlighting
Portfolio Website
Modern portfolio website with dashboard UI, dark mode, and smooth animations. Built to showcase projects and skills to potential employers.
Key Learnings:
- Mastered Next.js App Router and server components
- Implemented design systems with Tailwind CSS
- Created responsive layouts with modern CSS
Mita — Focus on What Matters
A minimalist task manager that limits you to three tasks per day to combat productivity anxiety. Features local SQLite storage, scheduled notifications, an Android home screen widget, streak tracking, and a glass-morphism dark UI. Offline-first and published on the Play Store.
Key Learnings:
- Built production Flutter app with native home-widget integration via platform channels
- Implemented exact-alarm notification scheduling with battery-optimization handling
- Designed a UX constraint (3-task cap) as the core product philosophy, not just a feature
Indian Road Driving Coach
AI-powered driving assessment that learns ideal driving behavior from YouTube videos of good Indian driving and scores user performance. Uses behavioral cloning with contrastive learning (good vs aggressive examples) and combines deviation from ideal predictions with anomaly detection for scoring.
Key Learnings:
- Trained a MobileNetV3 regression head to predict steering and throttle from video frames
- Combined behavioral cloning with contrastive learning for scoring, not just imitation
- Engineered a CPU-only training pipeline that runs within 8GB RAM constraints
VOC Gas Sensor Profiles Dataset
A synthetic dataset of 5,000 MOX sensor readings across 200 chemical compounds (alcohols, ketones, VOCs, combustion gases, etc.), each with a 10-step temperature-cycle resistance profile. Published to Hugging Face under Apache 2.0 for training gas classifiers as a baseline alternative to proprietary sensor data.
Key Learnings:
- Built a physically-grounded MOX resistance model (temperature dependence, power-law gas response, noise)
- Generated a reproducible dataset with a seedable generator script
- Distributed an open dataset on Hugging Face with a working classifier example
NanoAI — AI Image Prompt Gallery
A Pinterest-style mobile gallery for discovering and sharing AI image generation prompts. Masonry feed with infinite scroll, one-tap prompt copy, offline bookmarks, user submissions with an admin review queue, category filters, and AdMob monetization. Clean Architecture with Firebase backend.
Key Learnings:
- Implemented a masonry feed with infinite scroll and thumbnail lazy-loading via Coil
- Applied Clean Architecture (presentation → domain → data) in a community-driven app
- Designed Firestore security with authenticated-write content moderation
SAHE National Cadet Corps Unit Website
A multi-page static website for the 1(A) EME COY NCC Unit at VRSEC — home, unit info, events, gallery, and contact pages. Features an aim/motto popup, typed.js hero animation, and sections for leadership and unit activities.
Key Learnings:
- Built a polished multi-page static site with custom CSS and interactive elements
- Integrated typed.js for dynamic hero text and a lightbox-gallery experience
- Handled responsive navigation with active-page highlighting in vanilla JS
Vera Bot — AI Merchant Assistant
An AI-powered WhatsApp merchant assistant built for the magicpin AI Challenge. Uses a 4-context composition framework (category knowledge, merchant state, trigger events, customer context) to generate personalized business-growth messages, with trigger-based routing and multi-turn conversation handling via LLM.
Key Learnings:
- Designed a 4-context composition framework for personalized, trigger-relevant messaging
- Built WhatsApp-session-window-compatible output with multi-turn reply handling
- Routed distinct message templates (research digests, perf alerts, recall reminders) by trigger
IntelliCredit — AI Credit Decisioning
End-to-end AI-powered credit appraisal prototype for SME and corporate lending. Ingested documents, extracts financial ratios (DSCR, D/E, etc.), runs web research agents for due diligence, scores with primary insights, and auto-generates a full Credit Appraisal Memo with counterfactual explanations — cutting decision TAT from weeks to hours.
Key Learnings:
- Orchestrated an async 8-stage pipeline (ingesting → extracting → validating → researching → scoring → deciding → CAM)
- Blended deterministic banking covenants with LLM qualitative scoring and MCLR-linked pricing
- Generated explainable decisions with counterfactual narratives and document export
Heart Rate Monitor
Real-time heart rate monitoring system: ESP32 with KY-039 sensor streams samples over WebSockets to a FastAPI backend (SQLite-persisted) which fans out to a Flutter dashboard with live charts and device-key auth.
Key Learnings:
- Built a full IoT pipeline: hardware sensing → websocket ingest → persistence → live UI
- Implemented real-time fan-out with WebSocket subscriber management and device-key auth
- Integrated fl_chart for live vitals visualization in a mobile dashboard