Open-source distributed engines, reinforcement learning environments, low-latency telemetry pipelines, and tactile developer tools built for longevity, determinism, and software craft.
City-scale flood risk management simulator built to the OpenEnv architecture. Agents orchestrate barrier deployment, pump balancing, floodgate control, and civilian evacuations across a structured 8×8 grid under stochastic rainfall. Designed to evaluate LLM agents on spatial logistics with real-world consequence — flooding damage, resource exhaustion, and civilian risk. Meta PyTorch Hackathon Finalist.
AI-powered legal assistant built with a Flutter frontend and Python backend. Accepts uploaded legal PDFs and supports natural language queries, contract drafting, NDA generation, and clause classification across multiple API endpoints. Targets individuals, law firms, and legal departments needing accessible document intelligence. Course project in Software Engineering (CSBB 254) at NIT Delhi.
A cross-platform version control system written in C++ implementing the core workflows of Git — repository init, staging, committing, branching, merging, reverting, and log viewing. File data is compressed using Huffman coding and all repository state is stored in a hidden .vcs directory, with branch references managed as lightweight pointers into the commit DAG. Semester project for Design & Analysis of Algorithms (CSBB 202) at NIT Delhi.
Full-stack FuelEU Maritime compliance platform built on Hexagonal Architecture. The backend fully decouples domain entities and use cases from framework and database adapters, with Prisma + PostgreSQL as the outbound adapter. Features include GHG intensity computation, compliance balance banking, penalty calculation, route comparison, and multi-vessel pooling. The React frontend mirrors the same clean-architecture split across routes, comparisons, banking, and pooling pages.
End-to-end AI platform for legal contract drafting, compliance analysis, and risk assessment. A multi-agent FastAPI backend orchestrates specialized agents for clause classification, compliance checking, loophole detection, and contract generation. Retrieval-Augmented Generation over a Pinecone vector store gives the system access to thousands of legal reference documents. Secure document storage and usage analytics are handled via Supabase. Supports multiple legal jurisdictions including US and India.
Next.js 14 admin dashboard for content moderation, user management, and analytics. Built on Prisma ORM with a MySQL backend, it covers moderator search, post and user trend analytics, real-time database status monitoring, and a full seeding pipeline. Designed as a production-ready setup with a clean Tailwind UI. Course project in Database Management Systems (CSBB 204) at NIT Delhi.
Self-supervised contrastive learning system that trains a CNN encoder on augmented mel-spectrograms to produce audio embeddings without requiring labeled data. The learned representations support music genre classification, song recommendation, and audio similarity search. Preprocessing converts raw audio to both standard and Mel spectrograms; the encoder and projection head are trained with a SimCLR-style objective on augmented views of each track.
AI-powered healthcare facility recommendation system with full traceability. A Next.js frontend posts search queries to a FastAPI backend that runs a multi-step pipeline: facility search, data extraction, cross-verification across sources, trust scoring, and ranking. Results are surfaced with verifiable trust scores, distance metrics, and insurance compatibility, helping patients make informed care decisions.
Medical history record management system serving patients, doctors, and public health authorities from a single application. Supports diagnostic report ingestion, role-based access control across three user types, and epidemic hotspot detection through spatial clustering — surfacing disease concentration patterns to health authorities in real time.
Chat application with persistent history, bot commands, and an advanced multimedia pipeline. Messages are stored in SQLite and auto-restored on reconnect. Uploaded images are converted to WebP with live PSNR quality stats. Audio files are compressed via FFmpeg to AAC 64kbps and analysed across frequency bands — reporting compression ratio, spectral correlation, and bitrate metadata side-by-side with the original.
Three-tier career guidance system. A Python ML backend deployed on Vertex AI clusters user assessment data into career profiles. A Flask orchestration service on Google Cloud Run combines those clusters with Google Gemini API responses to produce personalised development roadmaps and career guidance. A React/Vite frontend handles user assessment input and conversational chat with the guidance engine.
Inbox management simulation environment built to the OpenEnv architecture. Agents process a realistic email stream — classifying messages into 8 categories, assigning priority levels, routing to the correct team, and reading threads when context requires it. Costs accumulate from SLA violations, wrong routing, and delayed action, creating a rich decision-making benchmark for language model agents operating under structured workflow constraints.
MLOps-style batch job in Python with a focus on reproducibility, observability, and robustness. Reads OHLCV market data, computes rolling averages on closing prices, generates binary buy/sell signals, and writes structured output to metrics.json and a run.log. Configuration is fully externalised via config.yaml, and the job handles malformed rows, empty files, and missing config gracefully. Docker-ready with no hardcoded paths.
End-to-end RAG application built with Next.js. The ingest phase chunks pasted text via LangChain, embeds it with Google Gemini, and stores vectors in Pinecone. At query time, the top-K candidates are retrieved from Pinecone, reranked by Cohere for precision, and passed to Gemini to synthesise a cited answer. Answers include inline citations linking back to the source chunks in the knowledge base.