the work

Case studies

Complete engineering records, not highlight reels — every eval published, every failure named, and the method shown wave by wave.

doc CS-2026/07 · rev C▶ real-data stage live 14B · Qwen3document extractioninsurance

Teaching a 14B model to read insurance forms — and prove it

Four ACORD form types, fully synthetic training data, frozen stress instruments, pre-registered gates — and a live real-data program tracked on the page. 0.986field acc · stress exam 40/40traps refused ~12 htotal GPU compute Read the record
doc CS-2026/06✔ shipped 4B · Qwen3legal reasoningproptech

Fine-tuning a 4B legal clause analyzer to 98.7%

A lease-clause verdict model that argues from retrieved law, knows when to ask a human, and cost under five dollars of GPU time to train. 98.7%verdict accuracy 53/53zero false accusations <$5total GPU spend Read the record
your model here The next record could be yours.

A 30-minute audit maps your use case to numbers like these — accuracy, latency, cost. Honest answer guaranteed, either way.

Book a free model audit no retainer · baseline guarantee