Engineering student in Artificial Intelligence & Data Science. I build agents, developer tools, and full-stack products — and I build them in public.
🌐 aman-on-bits.vercel.app — my site: writing, projects, and what I'm currently shipping.
- Agents & dev tools — coding harnesses, terminal UX, and the plumbing that makes models useful
- LLMs from scratch — pretraining, post-training, and evals on hardware I actually own
- Full-stack — real-time apps, self-hosted infrastructure, and native Linux tooling
| Project | What it is |
|---|---|
| kairo | Engineering intelligence over GitHub history — a temporal knowledge graph built from commits, PRs, and issues that reconstructs why a codebase looks the way it does. |
| sprout | A 32.5M-parameter story model trained from random initialization in seven hours on one RTX 4060. Weights on Hugging Face. |
| llm | A 101M-parameter LLM trained from scratch on one 8 GB laptop GPU — plus chessLLM, a 94M model that builds an internal board from move text alone (linearly probeable at 85.6%). |
| amzn-ml-2025-pricing | Amazon ML Challenge 2025: 4-bit LoRA on Qwen3.5-2B with a monotone quantile head for SMAPE-optimal price prediction, on a single laptop GPU. |
| chat-app | Real-time chat with rooms, DMs, OAuth, typing indicators, and full-text search. React 19 + Convex. |
| minihub | Self-hosted Git service in Go — HTTP/SSH repo hosting, pull requests, webhooks, and CI, with a React frontend. |
I write up what I learn, including the parts that didn't work: miniLLM on a laptop · the chess world model · a 214M training run that blew its quota
Site · X / Twitter · LinkedIn
82+ public repos and counting — always shipping something.