AI deployment at Pre6.ai. The work between a working model and a working company: what ships, for whom, in what order, and what has to be true before it does.
GitHub is a public reading list — small studies of the stack I help decide whether to put in production.
- mosaic — tokenization, the first cost and quality decision · demo
- nabla — autograd, so “backprop” is a picture · demo
- loom — a small GPT, enough to talk about the model without waving at a paper · demo
- warren — nearest-neighbor search, the retrieval step in every RAG deploy · demo
- agent-evals-lab — reliability, tool-use, latency, and cost as release criteria
- stencil — constrained decoding, when output must be valid structure · demo
- winnow — what to keep inside a token budget · demo
- mend — repairing malformed JSON from model output · playground
- semcache — semantic cache, so you do not pay twice for the same question
- rag-safety-gateway — injection, secrets, and PII as a deploy gate