Created a real-time collaborative IDE with Next.js, Yjs, Electron, WebSockets, Babel AST analysis, and multi-agent LLM workflows with Ghost Agents that augment development with live collaboration, predictive conflict resolution, and automated code visualization. Won the AIE Code Agents Hackathon.

Built 10,000 AI customer replicas to test email campaigns before sending. See who opens, clicks, converts—or unsubscribes—in under 800ms. Plus a Gemini-powered project simulator that shows ripple effects when timelines change.

A unified visual RAG system for security and law enforcement that processes video footage (body cams, surveillance, dashcams) with natural language search across visual content, spoken audio, and detected objects. Features multi-modal search, smart deduplication, real-time object detection, and click-to-track segmentation.

Built a real-time computer vision steering engine for Slowroads.io in under 60 minutes. Features smart frame skipping that decouples AI from render loop (2x FPS), custom PWM steering algorithm to fake analog input from binary keyboard, and optimized MediaPipe hand tracking with zero visual overhead.

A pixelated voxel building game with first-person controls, multiple block types, dynamic weather, save/load system, and particle effects. Built with React 19, Three.js, React Three Fiber, and Zustand for state management.

A minimal GPT-style language model built from scratch in PyTorch, trained on Shakespeare. Features a 4-layer transformer with 4 attention heads (3.2M parameters), character-level tokenization, causal self-attention, and top-k/nucleus sampling for text generation.

An intelligent negotiation system that simulates realistic car price negotiations between buyers and sellers. Features AI-powered dynamic response generation, real-time sentiment analysis, multiple negotiation strategies (Stand Firm, Split the Difference, Final Offer, Walk Away), and progress tracking with negotiation metrics.

A chat-based AI journal built on Cloudflare Workers using a custom agent pattern with Durable Objects for persistent state. Features a 5-entry context window for Llama 3.3 70B responses, automatic mood analysis via separate AI calls, SQL emulation layer for querying entries, and stateful memory across sessions—all running on the edge with zero cold starts.

Deployed an autonomous wildfire suppression system with multi-agent RL (Dueling DQN) drone swarms. Engineered a multi-objective reward system balancing suppression, efficiency, and safety, achieving 85% mission success and 2.4x more fires extinguished vs. rule-based drones. Awarded Best Presentation at the Qualcomm Snapdragon Multiverse Hackathon at Princeton.

AI tutor that helps students learn and understand complex topics through personalized learning paths and interactive quizzes. Learn from the best professors and teachers around the world.

AI tool to convert live vet consultations into prescriptions. Built a custom LLM that can understand the context of the conversation and generate prescriptions.

Leading development at Chewy of an intelligent system creating rich customer cohorts for precision marketing and campaign optimization.

Machine learning algorithms predict optimal crops based on soil conditions with 93% accuracy. Published research in IJRASET journal showcasing innovative agricultural AI solutions.

End-to-end health solution using MERN stack, serving 500+ users with personalized medical tracking, diet plans, and workout routines.

YOLOv7-powered attendance system reducing manual effort by 50% while serving 200+ employees with automated face recognition.