jev-311-heatmap
Research pipeline that scores NYC 311 complaint descriptions with Jev and maps severity and safety signals on a city grid.
Overview
NYC 311 JEV Heatmaps is a research demonstration from CompleteTech LLC AI Research. It downloads a bounded slice of NYC 311 service reports, evaluates each distinct complaint description with Jev, and maps the resulting signals on a projected 500-metre grid as interactive Folium HTML maps and print-ready PNGs. Layers include complaint count, scoring coverage, mean potential-impact severity on a Jev rubric, mean probability of a defined safety-hazard proposition, an answer-ambiguity proxy, and an illustrative composite.
For 1 September 2026 the README reports 9,903 reports downloaded, 9,698 mapped, 2,043 occupied cells and 634 distinct descriptions, with 205 excluded for invalid coordinates, 634 API calls without retries and 539,979 input tokens; repeated descriptions share one evaluation. Responses are validated and cached. The authors stress that the maps show reported complaints and model-estimated potential impact, not verified hazards or population-adjusted risk, and that blank cells are not evidence of no problem.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Score | What is the potential-impact severity of this reported condition? | — |
| Yes / no | Does this report describe a safety hazard? | yes · no |
How it works
- Download a bounded slice of NYC 311 reports
- Deduplicate descriptions and query Jev once per distinct text
- Validate and cache responses
- Aggregate per 500 m cell and render maps
Strengths
- Caching cut 9,698 reports to 634 API calls
- Clear methodology and run metadata
- Explicit caveats about what the maps do and do not show
Caveats
- Model-estimated impact, not verified hazard data
- Single-day demonstration
At a glance
- Type
- Service
- Status
- research
- Language
- Python
- Runtime
- CLI pipeline, Folium maps
- Has code
- yes
- Has tests
- yes
Use cases and integrations
How curated lists describe it
The live run excluded 205 reports with missing or invalid coordinates, completed \*\*634 API calls without retries\*\*, and reported \*\*539,979 input tokens\*\*. Repeated descriptions share one evaluation.
logicrw/awesome-jev-projects
Frequently asked questions
What does Jev score?
Each distinct 311 complaint description for potential-impact severity and whether it describes a safety hazard.
How many API calls did a day take?
634 calls for 9,698 mapped reports, because repeated descriptions share one evaluation.
Do the maps show real hazards?
No. They show reported complaints and model estimates, as the README warns.
Related projects
Mines timestamped driving scenes from vehicle logs by asking Jev lateral and longitudinal manoeuvre questions per moment.
ProjectbetaPythonFrench library ranking ADEME emission factors for an activity label, splitting code rules from Jev's semantic match.
ProjectprototypeJavaScriptBrowser-based systematic-review extraction tool where Jev points at the lines that answer each question and code copies verbatim quotes.
ProjectbetaJavaScriptNextflow plugin exposing Jev noul, choice and score judgments as pipeline functions you can gate on with ordinary operators.
PluginbetaGroovyRust research workbench where Jev gives a cancer-of-unknown-primary origin distribution over 22 classes, gated locally for review.
ProjectresearchRustJev-inspired PyTorch research framework that encodes an overnight sleep recording once and answers many Choice, Noul and Score questions.
ServiceresearchPython