Research note on how we tune local models for enrichment — and how those models show up in Orchestration Lab as our own provider inference endpoint.
What this covers
We tune and evaluate our own local models for enrichment jobs — classification, extraction, summarization, and signal cleanup that feed product pipelines. Work includes dataset curation, eval harnesses, quantization tradeoffs, and when a local model is the right tool versus a hosted API.
How it connects to Orchestration Lab
Tuned local models are not a separate product surface. They are registered as a custom provider in the AI Orchestrator pool (baseUrl, health check, request/response templates). The same generate() call can prefer local inference for cost or privacy, then fall back to OpenAI, Google, Anthropic, or other adapters when metrics say so.
See the Orchestration Lab project page for the full inference map: local + hosted providers, routing strategies, and the metrics we use to decide.
