A controlled ablation over what people assume about RAG. Chunking, retrieval and context assembly measured on one corpus with one eval, reporting recall alongside context cost so the trade is visible. Most benchmarks let a strategy win by quietly using more tokens.
911K Chicago traffic crash records end to end: wrangling, geospatial hotspot clustering, socioeconomic regressions, severity prediction. The interesting part wasn’t the model — it was finding which neighbourhood factors survived controlling for traffic volume.
A Chrome extension that reads a LinkedIn profile and generates DISC personality insights, running against a local Ollama model so nothing leaves the machine. Selenium handles the scraping; the prompting layer does the rest.
A custom Python autoscaler watching SQS queue depth to bring EC2 instances up and down, holding 1.5s latency under heavy load. Then migrated to Lambda — 20% faster at 1.2s per request, cheaper, and no autoscaler left to maintain.
Compared the computational efficiency of the two approaches on a language dataset, at the Confluence Lab.