InfraWatch

Satellite imagery in, SLA-tracked enforcement notice out. A vision-AI pipeline for flagging illegal construction in Indian municipal corporations, with confidence-based SLA routing, multi-provider AI failover, and a legal-notice generator constrained to a fixed statute list.

Google Solution Challenge 2026 project, team of two. Prototype, not a validated deployment.

---

Why

Bengaluru has an enormous stock of buildings and a small enforcement team, so detection-to-notice time is measured in weeks. InfraWatch explores closing that loop in software: a satellite tile goes to a vision model, a detection becomes a case with a response deadline, and the case gets a draft legal notice for an officer to review.

No pilot has been run, so there are no impact or accuracy claims here. Numbers that can be measured honestly are in eval/; vision accuracy is explicitly pending.

Features

Area	What the code does	Where
Vision analysis	Single-image and before/after analysis via Gemini; returns bounding boxes, violation type, confidence, risk level as JSON	server/services/visionAI.js
Confidence-based SLA	4h if confidence >= 90, 12h if >= 80, else 24h; status ON_TIME / AT_RISK (past 80% of SLA) / BREACHED	server/services/sla.js, server/routes/violations.js
AI failover	Gemini primary; on 429 / quota / 503 / timeout the same request is retried on Groq (Llama 4 Scout for images, Llama 3.3 70B for text)	visionAI.js callAI
Legal notices	Draft notices from a prompt that restricts citations to a fixed statute list and tells the model to use a fallback phrase when unsure	visionAI.js, gemini.js
Case workflow	Roles (admin, commissioner, inspector, field officer), ward-scoped access, notice templates, activity log of case actions	server/routes/*, server/middleware/*
Crisis feed	Dashboard banner polling every 15 s for high-confidence unresolved cases and SLA breaches	CrisisResponseBanner.jsx
Citizen portal	Public, no-login report endpoint, in-memory limit of 5 reports/minute per IP	server/routes/citizen.js
Other	Case-review dossier, chat over a violation, compliance report, analytics, PWA shell, /health page	server/routes/vision.js, analytics.js, client/

Architecture

flowchart LR
  subgraph Client["Client - Vercel"]
    UI["React 18 + Vite<br/>Leaflet map, Recharts, PWA"]
  end
  subgraph Server["Server - Render (Express)"]
    R["Routes<br/>violations / vision / notices / citizen"]
    AUTH["JWT auth + role and ward checks"]
    SLA["sla.js<br/>confidence to SLA tier"]
    AI["visionAI.js callAI<br/>timeout + failover"]
    DB[("sql.js SQLite<br/>auto-seeded on first boot")]
  end
  G["Gemini 2.5 Flash<br/>(primary)"]
  Q["Groq Llama 4 Scout / Llama 3.3 70B<br/>(fallback)"]
  UI -- "/api/* rewrite (vercel.json)" --> R
  R --> AUTH --> DB
  R --> SLA
  R --> AI
  AI -- "try first" --> G
  AI -. "429 / 503 / quota / timeout" .-> Q

Loading

Vercel rewrites /api/* to the Render backend (client/vercel.json); locally, Vite proxies /api to the server.

Design decisions

Confidence-based SLA tiers. Higher-confidence detections get shorter response deadlines, so scarce officer time goes to the likeliest true positives first. The mapping is a 10-line pure function (sla.js) so it is testable; boundary and monotonicity tests are in eval/. Known rough edges, documented rather than hidden: the scale is implicitly 0-100 (a 0-1 fraction lands in the 24h tier), and the crisis feed lists cases from 85 while the 4h tier starts at 90.

AI failover in one place. Every vision/text call in visionAI.js goes through a single dispatcher with a per-call timeout (default 30 s, AI_TIMEOUT_MS). Only retryable errors (429, quota, 503, unavailable, timeout) trigger failover; a 400 surfaces immediately instead of burning the second provider. Verified against stubbed providers in eval/tests/failover.test.js. Limitation: the notice route in gemini.js uses one provider and does not fail over.

Citation guardrail, two layers. The prompt names the only statutes and sections the model may cite (Karnataka Municipal Corporations Act 1976 s.308/321/321A/322, BBMP Building Bye-laws 2003, Karnataka Town and Country Planning Act 1961) and instructs a fallback phrase when unsure. Prompts alone cannot prevent hallucination, so server/services/citationGuard.js also checks every AI-drafted notice against the allowlist and replaces any off-allowlist citation with the fallback phrase, reporting what it removed in citationsRemoved. Measured before/after: the app passed 20/20 fabricated citations through with no guard, and catches 20/20 with it (0/20 false positives on valid text; synthetic fixtures, not a claim about real model hallucination rates). Every notice is still a draft for officer review, and the allowlist has not been verified against the statutes by a lawyer.

Run locally

Requires Node >= 18 (Render config pins 20.11.1). A free Gemini key is optional for browsing the seeded demo; AI calls need it.

git clone https://github.com/tchxm/Infrawatch.git && cd Infrawatch
npm install
(cd client && npm install)
(cd server && npm install)
cp server/.env.example server/.env      # add GEMINI_API_KEY (and optionally GROQ_API_KEY)
npm run dev                             # client :5173, server :3002

The database auto-seeds on first boot if empty: 214 synthetic violations across 15 wards, users, and notice templates. Seed login is printed by the seed script (server/db/seed.js). Run the eval with cd eval && npm test (after the server install).

Key env vars (server/.env.example): AI_PROVIDER, GEMINI_API_KEY, GEMINI_MODEL, GROQ_API_KEY, GROQ_MODEL, JWT_SECRET, PORT, AI_TIMEOUT_MS.

Screenshots

Dashboard + crisis banner	Map view	Vision analysis
TODO: docs/screenshots/dashboard.png	TODO: docs/screenshots/map.png	TODO: docs/screenshots/analysis.png

AI notice draft	Citizen report portal	Analytics
TODO: docs/screenshots/notice.png	TODO: docs/screenshots/citizen.png	TODO: docs/screenshots/analytics.png

Known limitations

• The demo's "city scan" and hotspot detections come from seeded canned results (server/routes/vision.js, getSeededDetection), so demo detections are not live model output. Non-hotspot images do go to the model.
• The single-image prompt tells the model to flag something whenever buildings are visible, which likely biases toward false positives. Vision accuracy is not measured (pending labeled data, see eval/).
• Officer assignment: the UI text mentions round-robin, but no round-robin logic exists in the server code; full-pipeline assigns new cases to the requesting user, and manual assignment is via the violations API.
• Model defaults are inconsistent: .env.example/render.yaml use gemini-2.5-flash, but gemini.js hardcodes gemini-2.0-flash for its text path and visionAI.js falls back to gemini-2.0-flash if GEMINI_MODEL is unset.
• After a failover, the _provider/_model metadata on parsed vision results still says Gemini.
• Data is synthetic; sql.js persists to a local file, so hosted state is ephemeral on free tiers.

My contributions

Mohammed Afnan (GitHub: tchxm).

What the git history supports (fork tchxm/Infrawatch of LordDevdeep/Infrawatch; 9 commits total, 2 author identities):

• 4 commits authored as Afnan <kariox364@gmail.com> on 2026-07-18, all touching only README.md (documentation rewrite and fixes).
• 5 commits authored by LordDevdeep (2026-04-20 to 2026-05-03) including both "first commit" commits that contain the entire codebase (server, client, AI services) and a README rewrite.
• Work added in this fork after that: the eval/ harness (SLA, failover, citation-guardrail and vision-eval tooling) and extraction of the SLA mapping into server/services/sla.js (behaviour-preserving).

What the history does not show. The upstream history contains no commits authored under the Afnan identity for server/ or the AI integration layer. Backend and AI-layer work (provider failover, notice generation, SLA routing) may have been written by Mohammed and committed under the upstream owner's account (the code arrives in "first commit" imports), pair-programmed, or written by the co-author. Git history cannot distinguish these, so this README does not claim authorship of specific modules. Mohammed should replace this paragraph with an accurate first-person account.

Evaluation

See eval/README.md for methodology and results: SLA routing (9/9 boundary cases, monotone over a 201-point sweep), failover (7/7 scenarios behave as designed with stubbed providers), citation guardrail (app-side catch 0/20 before the guard was added, 20/20 after; 0/20 false positives on synthetic fixtures), and vision accuracy (pending, harness and labeling template included).

License

MIT, see LICENSE. Upstream: LordDevdeep/Infrawatch.