Multi-tenant platform (FastAPI, React, PostgreSQL) measuring how often brands get recommended across 7 AI shopping assistants, with daily scans for 25+ brands. Includes a fine-tuned ModernBERT cross-encoder (NDCG@10 0.98) driving edit prioritization, a LangGraph harness powering 4 tool-dispatching agents, and a hand-rolled MCP server for ops agents.

Authored a preprint on a decentralized, probabilistic AI memory infrastructure running as an MCP server — a cascading 3-layer retrieval model with client-agnostic hardware anchoring. Engineered an identity resolution engine using int8 quantization and Gaussian boosting (F1 0.82), eliminating context decay with a 41% retention boost at sub-4ms latency and recovering 94.5% of memory across disjoint apps.

Infrastructure converting store inventories into AI-accessible APIs — a plug-and-play system for Shopify/WooCommerce that instantly spins up MCP servers for live catalog access, plus an injection engine that pushes MCP adapters directly into GitHub repositories for custom sites, enabling natural language product querying, dynamic price sorting and automated checkout prep via AI agents. Built at the Entrepreneurs First Hackathon.

An AI-native multi-agent system on LangGraph that deciphers historical languages and ciphers via collaborative agents (Pattern Analyzer, Hypothesis Generator, Decipherer) communicating over A2A, with a compiled DSPy "Oracle" for coherence scoring.

Built an ML pipeline using LSTM + XGBoost to predict daily SOFR fixes, achieving a high R2 score of 0.97.

Led solo development of a video learning app and launched it on both the Google Play Store and Apple App Store.

Developed a full-stack MERN application to optimize appointment scheduling and manage patient queues.

Translated detailed Figma designs into precise, high-quality mobile app screens and components using Flutter.

An open-source Flutter agent stack that runs the agent inside the app rather than on a Python backend — ReAct loops, 7 model providers, per-call routing and on-device inference. Five published packages spanning the core runtime, stateful agent graphs, semantic memory, device tools and a drop-in chat UI, four of them at a perfect 160/160 pub.dev score. Powers genesis AI, which builds a working on-device agent from a one-sentence prompt.

Gamified Flutter and Firebase learning app for the Department of School and Mass Education, Odisha, with 1000+ questions across 20+ categories, AI-powered multilingual text and animated explanations for wrong answers, and a user study framework for a pilot with underprivileged children.

ML pipeline using DSPy and MCP to predict loop termination and non-termination (invariants and recurrent sets), with GCC & CBMC integrated into an 8-split workflow to verify every predicted property. Bootstrapped 10x more data via BootstrapFewShot and CBMC feedback loops, reaching 75% prediction accuracy.

On-premise solution using Llama3 and Meta LLMs via Ollama with RAG to match mobile device vulnerabilities against CWE entries for the DRDO, integrating Hugging Face for extended LLM support and achieving 65% matching accuracy.

Multi-node NLP pipeline in LangGraph formalizing natural language design requirements into DSLs for Wabtec Corporation, with a hallucination detection model built on INCOSE sentence generation and semantic similarity scoring, combined across models for robust validation.

Flutter app for the Government of India's Jal Jeevan Mission tracking water leakage across rural locations, with real-time leakage zones on a map and a complaint management system, backed by a Go API layer and PostgreSQL — set to impact 19Cr+ households on deployment to the JJM portal.

Flutter app for India's first AI self-checkout kiosk, allowing real-time item scanning and integrated payments.

On-prem workspace controller in Go orchestrating Docker containers with authentication and server redirection, Rocket.Chat behind Nginx, a GitHub-profile-based AI task matching recommender, semantic document retrieval for a documentation explainer, and a Selenium automation tool — 5th place at Inter IIT Tech Meet 12.0.

1st Runner Up at SnT Hackathon 2023. An Android app using QR codes and Firebase to manage campus entry-exit.

1st Runner Up at Techkriti. A MERN stack website with 3D image viewing for properties and precise map locations.

1st Place at OOSC'2024. Automated extraction and analysis of embedded web links using Selenium scraping and the Gemini API to generate factual questions, with TFIDF vectorisation for relevance mapping — achieving 83% accuracy.