ML / Bio / Research — I build the machinery underneath the tools, instead of just calling them.
This year moved in one direction: from using ML systems to implementing the machinery underneath them.
Software engineering → ML → mathematical foundations → computational biology
→ database internals → geometric diffusion → LLM internals → AI infra
Each stage was chosen to force the next question down a level — how attention actually works as tensor contractions, how a database persists a B+ tree to disk, how a diffusion model has to be reparameterized to run on a manifold like SO(3) instead of flat Euclidean space.
- fenstersql — a relational database engine from scratch in C: SQL parsing (Flex/Bison), page-based storage, B+ tree indexing, catalogs, disk persistence. Built to actually understand what a database is doing, not just to query one.
- protdiffusion — a research-oriented diffusion model over protein backbone geometry. Represents structure as SE(3) rigid frames, implements Invariant Point Attention, and works through the SO(3) rotation-diffusion problem directly. Started as a from-scratch reconstruction of RFdiffusion's ideas, not a reproduction of it.
- scgpt-mini — an implementation of the scGPT paper for single-cell foundation models.
- alphafold2-mini — an implementation of the AlphaFold2 folding transformer.
- bio-agent — a biology-focused, tool-calling agent, SFT/LoRA fine-tuned on an open-source LLM.
- breast-cancer-prognosis — prognosis prediction from TCGA transcriptomic and clinical data, including catching and correcting a data-leakage issue that was quietly inflating model accuracy.
- LLM inference and serving — vLLM, GPU/VRAM sizing, throughput/latency tradeoffs, scaling and routing strategies
- AI agent architecture — tool boundaries, idempotency, retry and failure handling
- Kubernetes / AI infrastructure (CKA/CKS)
- Robotics and reinforcement learning as a next direction