01
Models are commodities. The product is the system around them — retrieval that grounds, evals that verify, interfaces that earn trust.
I build AI the way systems people build engines: measured, evaluated, honest about failure modes. A pipeline is only done when it behaves under real inputs — not demo inputs.
That means grounding every answer in retrievable sources, testing with evals instead of vibes, and designing the human layer so uncertainty is visible, not hidden.
An ungrounded answer is a liability. Retrieval, citations, sources attached — always show the work.
If it isn’t measured it isn’t shipped. Datasets, regression suites, and honest error budgets.
The model is 20% of the work. Latency, streaming, state, edge cases — that’s where trust is built.
02
Shipped builds — live, working, and open to explore
01 / 2026
Ask where any satellite is in plain English and get live orbital answers. Two agents split the job — one talks, one does the math — with an MCP server as the exclusive gateway to N2YO, Celestrak, and Nominatim, plus true 3D slant-range distance.
02 / 2026
An F1 chatbot that routes every question to the right brain — RAG over 1,469 Wikipedia chunks for trivia, Text-to-SQL for stats — with hybrid BM25 + pgvector search, prompt-injection detection, and a live React chat UI.
03 / 2026
A secure agentic RAG assistant for your own documents. Upload PDFs and ask — a 3-stage hybrid filter retrieves, then a LangGraph loop critiques and reformulates until the answer is grounded or safely refused.
04 / 2026
Qwen3-8B fine-tuned on 10,656 F1 question-answer pairs — race results, qualifying, driver bios, champions. QLoRA on a Colab T4, weights published on Hugging Face, training notebook in the repo.
03
Multi-agent orchestration & production LLM serving
04
Source code, datasets, and weights — open for the community
Multi-agent satellite tracking — ADK agents, MCP gateway, real-time orbital mechanics.
F1 chatbot — dual RAG + Text-to-SQL pipelines with query classification.
Secure agentic RAG assistant — 3-stage hybrid filter, LangGraph self-correction.
Qwen3-8B fine-tuned on 10,656 F1 QA pairs — QLoRA via Unsloth, weights on Hugging Face.
05
Let’s build something intelligent.
Open to collaborations, open-source work, and interesting side projects. If you are building with AI — models, retrieval, agents, or anything in between — let’s talk.
Usually replies within 24 h