About this prototype

Rule-based and fully client-side; nothing on this page is transmitted anywhere. It demonstrates the intake-to-downstream pattern: extraction with per-field confidence, policy screening, human-in-the-loop exceptions, and an attributable audit trail. The production version is an agentic extraction pipeline with an evaluation harness, model-confidence routing, and posting into the systems of record.

• Extraction with evals, not vibesLLM extraction scored against a labeled regression set per document type; confidence thresholds tuned to route uncertainty to humans, and every model change re-scored before deploy.
• Human-in-the-loop by designApprovers see the finding, the evidence, and the suggested action in one place; decisions feed back into reusable rule and prompt patterns.
• Enterprise fitDeploys inside the firm's environment with observability, access controls, and a full audit trail; posts to the ERP and approval systems of record.

More working systems by the same builder

• russh.workThe full portfolio: six production AI systems, built and operated end to end.
• Realization GuardDocument-compliance agent prototype, scoped and shipped solo in one day with Claude Code.
• TrialEdgeProduction ML pipeline: LightGBM plus LLM document analysis, ingestion to inference.
• ODI Bot19-pass multi-provider LLM orchestration with per-pass contracts and citation enforcement.

Built as a single deployable file by Russell Harris · russh.work · [email protected] · 302-824-0225