TARCJA: judicial document extraction & dashboard

• Applied the multiagent document taxonomist to 5200+ judicial appeal PDFs — no per-document parsers.
• Designed a domain schema for judicial appeals: facts, actors, jurisdiction, lifecycle, risk, and outcome fields.
• Processed at scale: 95–100% core field coverage, 16.8s avg latency per document, p95 26.1s.
• Built a static Plotly + Jinja2 dashboard deployed on GitHub Pages with multi-run comparison view.

• Domain schema design for unstructured legal text: universal facets (actors, jurisdiction, lifecycle, risk) transfer across document types.
• Static dashboard from JSONL: Plotly + Jinja2 generates standalone HTML — no server needed post-extraction.
• Multi-run comparison UI: version selector and compare view expose schema or data changes across extraction runs.
• Local-first tools for iteration: run schema induction and extraction locally before deploying to cloud workers.

Started as a submission to the Gemini API Developer Competition on Devpost.

The domain: a corpus of judicial appeal documents in unstructured PDF format, with no consistent layout or field structure.

Goal: extract structured, queryable records from each document automatically, then explore patterns across the full corpus.

• Reused the multiagent document taxonomist as the extraction backend — no need to rebuild schema induction or cloud workers for this domain.
• Designed a judicial-specific schema using the SchemaDesigner agent on a sample of documents: key fields include parties, jurisdiction, proceeding type, lifecycle stage, and outcome.
• Chose static HTML output (Plotly + Jinja2) for the dashboard: no database or server needed to share results, and GitHub Pages is free.
• Multi-run structure in the dashboard: each extraction run produces its own view, with a comparison page to surface changes across schema versions or data updates.

• Extraction backend: multiagent document taxonomist platform (FastAPI + Cloud Functions + Firestore).
• Schema: induced via SchemaDesigner agent on a sample of judicial PDFs, then refined manually for domain accuracy.
• Output: JSONL per run, stored locally and fed into dashboard_tarc for visualization.
• dashboard_tarc CLI: builds standalone HTML dashboards from JSONL + schema manifest — one command, no server.
• GitHub Pages deployment: static dist/ pushed to the docs/ branch, no CI/CD needed.

• 5200+ judicial documents processed end-to-end.
• 95–100% field coverage on core schema fields (parties, jurisdiction, proceeding type, outcome).
• Avg LLM extraction latency: 16.8s per document; p95: 26.1s.
• Static dashboard deployed on GitHub Pages with multi-run comparison and run selector.
• Continued development after the hackathon: schema versioning, additional facets, observability improvements.

• Dashboard (GitHub Pages)
• GitHub repository