Know which door to knock on first.
View the one-pager · Try the live AWS demo
Childhood lead-exposure risk triage agent for the Mireye Build Challenge.
Lead poisoning in children is almost entirely preventable, but by the time it is caught the damage is done. CDC estimates roughly 500,000 U.S. children have a blood lead level at or above the 3.5 μg/dL reference value. Health departments have the budget to test, but not the budget to decide which addresses to visit first. LeadSentry does that triage: it reasons over physical-world data per address, scores risk, and drafts the follow-up action (outreach letter or test-kit dispatch), with every number cited to its source.
Mireye's catalog does not include a year-built field, so LeadSentry pairs Mireye with US Census ACS Table B25034 (Year Structure Built) at the tract level, joined through Mireye's tract_geoid field. Pre-1980 housing share is the backbone of CDC/HUD lead-risk models.
| Score component | Weight | Source |
|---|---|---|
| Pre-1980 housing share | 0-50 | Census ACS B25034 (via free Census API key) |
| Legacy contamination (Superfund / brownfield / RCRA TSD / open LUST) | 0-30 | Mireye (EPA SEMS / ACRES / RCRA / UST Finder) |
| Water service gap (outside CWS area, high domestic-well density) | 0-20 | Mireye (EPA CWS Service Areas V3.0) |
Bands: 0-30 low, 31-60 moderate, 61-100 priority.
A vendor-agnostic LLM (any provider; set LLM_MODEL=provider:model) drives the loop: fetch Mireye facts, join Census tract data, compute the deterministic score, reason about data quality (parcelGrade=false caps confidence; heavy nulls downgrade it), then decide and record an action. The action writes a real artifact to actions/.
Every report also shows an Agent vs. rule baseline table. The same rule-based policy used by the offline fallback is applied to the same data, so you can see exactly where and why the agent upgraded or downgraded an action.
If no LLM key is available, the pipeline falls back to clearly-labeled, rule-based offline triage and still produces a complete report.
npm install
cp .env.example .env
# Add MIREYE_API_TOKEN, CENSUS_API_KEY, and LLM_MODEL
npm start -- --input data/sample-addresses.json --output report.mdTokenless demo against clearly-labeled illustrative fixtures:
npm run demoThe SAM application exposes POST /score, runs the existing deterministic LeadSentry scoring pipeline, and stores each assessment in DynamoDB. The deployed Function URL defaults to demo mode so it can be evaluated without API credentials.
Start DynamoDB Local, create its table, build the Lambda, and start the local API:
docker run --rm -d --name leadsentry-dynamodb -p 8000:8000 amazon/dynamodb-local
npm run dynamodb:init
npm run sam:build
npm run sam:localIn another terminal, score one of the illustrative demo addresses:
curl -X POST http://127.0.0.1:3000/score \
-H "Content-Type: application/json" \
-d '{"address":"436 Massachusetts Ave, Buffalo, NY 14213"}'Invoke the Lambda event directly when the HTTP layer is not needed:
sam local invoke LeadSentryFunction -e events/sample.json --env-vars env.local.jsonDeploy the public Function URL and DynamoDB table for the challenge:
npm run sam:build
sam deploy --guidedThe deployment outputs FunctionUrl, HttpApiUrl, and TableName. FunctionUrl is the public endpoint to link from the Builder Center article. AuthType: NONE is intentional for the public demo; remove the stack after judging or add authentication before using it with non-demo data.
On Windows and macOS, host.docker.internal lets SAM's Lambda container reach DynamoDB Local. On Linux, replace that endpoint in env.local.json with the Docker host address available to the SAM container.
Per-address triage across a whole ZIP is wasteful. ZIP mode screens an entire ZIP cheaply first:
npm start -- --zip 14213Stage 1 maps the ZIP to its census tracts (free Census ZCTA-tract relationship file and Gazetteer internal points, downloaded once and cached), scores each tract with one Mireye point sample and one Census call, then the agent turns the ranking into a canvassing plan: which tracts get door-to-door saturation, which get mailers, which get monitored, and where the expensive per-address stage-2 budget should go.
Add --deep N to run stage 2 automatically:
npm start -- --zip 14213 --deep 5LeadSentry selects the highest-priority canvass tract(s), samples N real addresses inside them (Census TIGERweb tract polygon plus OpenStreetMap addresses, both free and no keys), and runs the per-address agent. One command produces both the ZIP canvassing plan and the ranked per-address report with outreach letters and dispatch CSV.
If the ZIP screens mostly low-risk, --deep skips stage 2 and records that decision in the report so you can reallocate the budget elsewhere.
For public-facing reports, add --redact to replace street numbers and coordinates with placeholders:
npm start -- --zip 14213 --deep 5 --redactLeadSentry correlates its ZIP-mode scores with real childhood blood-lead outcomes from the NYSDOH Childhood Blood Lead Testing and Elevated Incidence by Zip Code dataset.
Across 26 ZIP codes in Erie County, NY, LeadSentry's land-weighted mean risk score shows a Spearman ρ of 0.55 and a Pearson r of 0.44 with the observed elevated-BLL rate per 1,000 children tested.
High-impact ZIPs align with the model:
- 14212: LeadSentry 70 · NYSDOH 89 EBLLs per 1,000 tested
- 14213: LeadSentry 66 · NYSDOH 87 EBLLs per 1,000 tested
- 14204: LeadSentry 57 · NYSDOH 57 EBLLs per 1,000 tested
Low-risk ZIPs also align:
- 14221: LeadSentry 43 · NYSDOH 0 EBLLs per 1,000 tested
- 14228: LeadSentry 35 · NYSDOH 0 EBLLs per 1,000 tested
Validate a single ZIP:
npm start -- --zip 14213 --validateRun the multi-ZIP study:
npm run validate:erieLLM_MODEL uses provider:model syntax and works with any provider the Vercel AI SDK supports:
openai:gpt-4o(needsOPENAI_API_KEY)anthropic:claude-sonnet-4-5(needsANTHROPIC_API_KEY)google:gemini-2.5-flash(needsGOOGLE_GENERATIVE_AI_API_KEY)deepseek:deepseek-chatordeepseek:deepseek-reasoner(needsDEEPSEEK_API_KEY)meta:muse-spark-1.2(needsMODEL_API_KEY)openai-compatible:llama3.1withLLM_BASE_URL(Ollama, vLLM, LM Studio, etc.)
Mireye is reached through its hosted MCP server (api.mireye.com/mcp) with a bearer token. If MCP auth is refused, LeadSentry falls back to the deterministic REST /v1/fetch automatically; same tools, same agent loop.
npm run typecheck
npm test- ACS is tract-level, not parcel-level. A v2 would weight by block group.
- The contamination sub-score is a distance-threshold rule. A v2 would fit it against blood-lead surveillance data per tract.
- Water-service fields flag regulatory gaps (private wells), not confirmed lead service lines. EPA's Lead Service Line Inventory would sharpen this.
Built by NaveenKumar Namachivayam for the Mireye Build Challenge.