I design and lead teams that turn large language models into real products — agentic RAG assistants, LLM fine-tuning pipelines, and computer-vision systems. 8+ years deep in Data Science, NLP, and Computer Vision, and I still love to build with my own hands.
I'm a Data Science Lead in Gurgaon who's spent the last 8+ years taking AI from idea to production — owning data strategy, architecture, deployment, and the team. I care about clean engineering, pragmatic system design, and building things people actually use.
Side projects I designed and shipped end-to-end — all open source, live, and documented.
An autonomous multi-agent simulation set in modern Gurgaon. Ten AI characters with distinct personalities live, work, and interact entirely on their own — each powered by an LLM, reading its own "soul file", keeping a private diary, forming opinions, and making decisions based on hunger, energy, mood, and goals. No player control — just watch the drama unfold.
A plugin that turns any folder of HTML into a Figma-style review canvas. Highlight text, click elements, leave comments, hit ▶ Process — Claude edits the source HTML, the page reloads with the changes highlighted, and you revert anything you don't like with one click. A fast, visual, reversible review loop that stays in the browser.
A beautiful cross-platform Markdown editor with first-class AI assistance and pixel-perfect PDF export. Multi-tab editing, side-by-side preview, a full formatting toolbar, on-demand AI continuation and refine tools over Anthropic / OpenAI / Google / Ollama, and OS-keychain-encrypted key storage. Privacy-first and polished.
Flagship AI systems I've architected and led with my team. Details generalised to respect confidentiality.
A multi-agent assistant that gives sales agents real-time answers from product knowledge bases — retrieval, reasoning, and response in one pipeline, deployed across retail and telesales channels.
A supervisor-based multi-agent chatbot replacing a legacy intent system — handling sales, claims, renewals, and service requests with guardrails, and voice support in a later phase.
A serverless computer-vision pipeline that detects mobile-device defects (screen, back, sides) from video using YOLO v8 and LLM-based OCR, built for customer onboarding.
A multi-stage data-preparation pipeline feeding a domain-specific LLM, fine-tuned with QLoRA 4-bit quantisation using SFT and CPT techniques.
What I reach for, organised by area.
Open to interesting conversations around GenAI, Agentic AI, and building AI teams. The fastest way to reach me is email — or connect on LinkedIn.