Field notes
Each note follows one question through a project: what I tried, what failed, what changed, and the evidence.
- Retrieval, Why my RAG pipeline needed BM25Four papers on the same topic sat on top of each other in embedding space. Keyword search found what vectors ranked #250.
- Reinforcement learning, What reward hacking taught me about RLA bonus for staying upright and slow taught a PPO lander to hover until time ran out: 0 landings out of 50.
- Systems, How I designed ownership rules for a WhatsApp → PMT syncSyncing support tickets between two systems turned out to be an ownership problem, not an API problem.
- Computer vision, Why I don’t publish an accuracy number yetThe floor-plan model at AHA is still training. Here is what exists instead: a labelled set of 4,035 images.