Abdul KalamUrbanist & geospatial analyst
Planned2026

Jurisdiction Intelligence OS

Permit-cycle benchmarks for 6+ jurisdictions, with an LLM extractor gated on a scored eval.

Not built yet. A committed build guide and product spec exist; no code has been written. The targets below are acceptance criteria the project must hit before anything publishes — not results it has achieved.

  • llm-agents
  • data-pipeline
  • planning
  • policy
  • python
  • 6+Benchmark jurisdictions targetedscope, not a result
  • 100+Board items to extractscope, not a result
  • ≥ 0.90Required precision on a 30-item gold setscope, not a result

Figures marked scope describe the size of the problem this project is specified to handle. They are not measurements — nothing has been run yet.

What this is

A working demo of a B2B product: jurisdiction scorecards benchmarking real permit cycle times from open municipal records, a project-fit checker driving a dynamic submission checklist, and a signal feed that extracts planning-board actions from New Jersey meeting minutes. Presented as a product — landing page, positioning, an explicitly illustrative pricing page — while every number in it comes from public records.

It complements the flood work by pointing the same data-engineering habits at a private-sector question, and by putting product thinking on the page alongside the pipeline.

Batch pipeline Python, weekly via Actions Permit ETL Socrata/CKAN open-data portals LLM extraction board minutes → structured actions Gold-set eval gates publication on measured precision Working store DuckDB benchmarks, crosswalks, rule engine inputs Committed artifacts each file ≤ 500 KB scorecards/*.json playbooks/*.json signals.json · alerts.json · meta.json Static app no server, no runtime API calls Vite + React + Tailwind Cloudflare Pages
How does a pile of municipal PDFs become a jurisdiction scorecard with no backend?Jurisdiction Intelligence OS build guide §6.1–§6.3.

The part that actually matters

The centrepiece is not the LLM extractor; it is the evaluation set that gates it. Thirty hand-labeled board items, precision measured per field, and a publication gate at 0.90 — an extraction pipeline without a scored gold set is a demo, not an engineering artifact. Every scorecard figure traces back to a recorded source extract, and every checklist item carries a citation to an official source.

Collection ethics are written into the guide as non-negotiable: honour robots.txt, identify with a contact email in the user agent, at least five seconds between requests to a host, cache everything so nothing unchanged is ever refetched, public documents only, and stop entirely on a 403 or 429. Published extractions reference the board, not named private individuals.

The spec’s original plan was to buy commercial permit data. This build deliberately does not — the spec itself argues the moat is the normalisation and extraction layer, so that is the layer being demonstrated.

Share this page

https://abdulkalam.pages.dev/projects/jurisdiction-intelligence-os/

Scan to open this page. Download the SVG for print.