Most of our work sits under NDA, so what follows is anonymised — the shape of the problem, the layer it lived in, and what moved. Detailed case studies are available on request.
A research desk had a document assistant that answered fluently and cited badly. Confidence was high, grounding was not, and compliance had stopped the rollout.
Rebuilt retrieval as hybrid search with a cross-encoder reranker, rewrote chunking around document structure, and put a grounding evaluation in CI so a regression fails the build.
Inference spend was growing faster than usage. The product team had optimised prompts twice and run out of surface-level ideas.
Moved the hot path to a distilled open model, applied 8-bit quantisation and continuous batching, and routed only the hard tail to a frontier API.
Clinical data could not cross the network boundary, which ruled out every hosted option the team had evaluated.
Designed and deployed an on-prem stack on existing GPU capacity — fine-tuned small models, private serving layer, full audit trail — sized against real utilisation rather than a vendor's quote.
We don't publish client names or screenshots. If you want the specifics — architecture diagrams, benchmark methodology, what went wrong on the way — ask, and we'll walk you through it.
Tell us which of these looks closest to your problem and we'll bring the detailed version to the first call.
Case studies shared under mutual NDA.