Dutch Forensic Psychiatric Report Quality Assessment
A privacy-first, fully local AI tool that evaluates Dutch forensic psychiatric reports against formal quality guidelines.
Free 30-min scope call · reply within 24h
Workflow & Architecture
How this system works
PDF/DOCX forensic reports loaded and parsed entirely on local hardware.
Report sections mapped to formal quality criteria and chapter labels.
Lightweight local model generates evidence-backed feedback with quotes.
Annotators refine outputs via Prodigy JSONL for iterative accuracy gains.
Overview
Built a local-only pipeline using lightweight local LLMs to ingest PDF/DOCX reports, map content to guideline labels, and generate evidence-backed feedback with direct quotes and section references — no data leaves the local environment.
The Problem
Compliance-sensitive forensic reports can't be sent to the cloud, but still need consistent, evidence-based quality review against formal guidelines.
The Solution
Delivered a local RAG pipeline with chapter-to-label mapping, evidence highlighting, and a human-in-the-loop annotation workflow for iterative accuracy improvement.
Key Features
- 100% local processing — no cloud data transfer
- Chapter-to-label mapping for deterministic, explainable evaluation
- Evidence highlighting with direct quotes and section references per criterion
- Human-in-the-loop annotation workflow (Prodigy JSONL) for accuracy refinement
- PDF/DOCX ingestion with structured report feedback
Outcomes & Business Value
- Delivered and live in production, running the full review workflow entirely on-prem for the client
- Annotation-based evaluation loop lets the client keep improving accuracy over time, without sending anything off-site
- Built for compliance-sensitive legal and forensic environments where cloud processing isn't an option
