Internal product · Knowledge systems
Coding Agent Memory Harness
Hybrid retrieval with a fine-tuned cross-encoder: worth 10× the latency.
Read →Lab
We build to understand. Experiments, internal products, and research from the problems we hit while building AI systems - labelled for what they are. No invented case studies.
Experiment · Models
Policy Classifier Agent
Prompt-defined policies, enforced in parallel by a fine-tuned 4B - before the frontier model ever answers.
Read →
Experiment · Applied ML
Cinema Screening Planner
A LightGBM ranker hit ~95% of this data's measurable ceiling - and proved the human scheduler was already good.
Read →
Internal product · Agents
RAG demos assume everyone can see everything. Companies don't work that way.
A team knowledge system built to scale to millions of documents - where permissions shape retrieval itself.
Read →