a small neural net learning to untangle two spirals, trained from scratch in numpy (pure SVG + CSS, no JS)
- ShilmanDB: a relational database engine in C++ built from scratch (SQL parser, B+ tree index, Volcano-style executor, buffer pool). Its learned join-order optimizer hits 94.1% exact-match accuracy, and a learned eviction policy beats LRU by 3.64% hit rate.
- text2sql-router small LLMs are cheap but often write wrong SQL. A 7B model answers every question first, its SQL is run, and a classifier trained on what happened (did it execute, what came back, token confidence, agreement across retries) decides which questions to escalate to a 32B. It matches the 32B's accuracy while paying for it on only 30% of questions (29% of the API cost), and on 68 databases it never trained on, it scores 2.4 points above the 32B while escalating just 20%.
- NN-Steiner reproduction: reproduced an AAAI 2024 paper where a neural network guides rectilinear Steiner tree construction for chip wire routing. Found the commented-out step that kept the authors' own model from matching their paper, then trained our own model that ties theirs (2.73% vs 2.65% above optimal) using about 1/36 of the paper's training. Solving each problem in its 8 rotations and mirror images and keeping the shortest tree cuts our error to 1.76% with no retraining (theirs: 1.87%).
- LLM agents in production: agent pipelines processing 20k+ signals a week, every prompt change gated by an eval harness (golden-file regression tests plus an LLM-as-judge at 92% agreement), with 66% lower cost per run from prompt caching and token budgets.