Research.
We publish notes on training, evaluation, and deployment for specialist models that improve inside real production workflows.
BYO SWE-grep: automatically train blazing-fast search sub-agents on your knowledge base
Search sub-agents that learn the structure of your internal knowledge for fast, reliable retrieval.
Purpose-built LLMs for dental note-taking
Frontier-style reasoning performance at a fraction of the latency.
Lumina: building self-improving evaluation through customer-in-the-loop refinement
An adaptive evaluation engine that learns to judge like a subject matter expert.
Attention-based attribution: what your model is actually looking at
Interpretable signals for understanding model behavior.
Upweight the strategy, not the tokens
Rationale-guided training teaches the why, not just the what.
Robust, sample-efficient SFT with prompt mutations
Low-divergence prompt variation for better performance at lower cost.
Training loss predicts evaluation performance
A practical evaluation signal for non-verifiable tasks.