A real-time video analytics pipeline that refuses to store unknown faces
Detect, track, check liveness, embed, match. One worker per camera, a SQLite source of truth, and a vector index that heals itself when the model changes.
7 articles in this category
Detect, track, check liveness, embed, match. One worker per camera, a SQLite source of truth, and a vector index that heals itself when the model changes.
An OpenAI-compatible endpoint, guardrails, a model registry, drift detection, and Prometheus metrics, packaged as a single Docker container.
Four public sources, a versioned lake on Hugging Face, dbt transforms, an IsolationForest, and a dashboard at the edge. The compute bill is zero.
Provider-independent, zero telemetry, and it writes your session to a local JSONL file. The constraints came first and the features followed.
A 156,000-word dictionary that runs in the browser, a router that only reaches for a model when you bring your own key, and no database on the other side.
An agent that runs long enough will forget what it decided. I implemented the GCC framework, which treats agent memory like Git, in seven languages against one on-disk format.
ART splits reasoning into six stages with explicit uncertainty and refuses when evidence runs out. GSM8K 55.6 percent at 0.29x cost in our runs.