- qwen3-8b
- time-entry extract · gate QA
- single node
- llama-8b-2stage
- narrative rewrite + revision
- pipeline-parallel × 2 AI PCs
- compliance engine
- UTBMS + privilege leak-check
- deterministic tool
legal
Billing narratives
Raw, cryptic attorney time entries in; client-ready, UTBMS-compliant billing narratives out. A deterministic compliance engine (not the model) flags invalid codes, block-billing, and vague language, and guarantees that no flagged privileged or confidential term survives in the client-facing text: it re-scans the rewrite and any survivor fails the gate closed. The on-prem story: attorney work product (client names, settlement figures, litigation strategy) never leaves the firm.
pre-warming the mesh (qwen + 2-stage chain)…
The case