CopperNick Vision

An AI-native cyclone impact and infrastructure vulnerability forecaster, paired with a passive sky-scanning early-warning subsystem. Built for Build with AI: Code for Communities (hack2skill / Google Cloud, Track 05, deadline 2026-09-30).

Live: https://coppernick-222356459208.asia-south1.run.app

About the name and GOSIM

This started as "CopperNick," scoped from a one-line brief ("a sky patrolling agent observer, Gosim (China) Vision OS hackathon") before any of GOSIM's actual event pages were checked. Once verified, GOSIM Shenzhen 2026 has no "Vision OS" hackathon or track — the real hackathon-shaped events there are Spotlight Shenzhen 2026 (an open call for AI-native products) and the OAIC Agentic Factory Hackathon (registration not yet open, as of this writing). "CopperNick Vision" is the current name, reflecting the vision/sky-scanning capability actually built — not a claim of GOSIM eligibility, which remains unconfirmed and secondary to the hack2skill submission this repo is actually built for.

Concept

Track 05's brief: build an AI-powered predictive risk and vulnerability modeling platform that simulates cyclone storm surges, predicts local rainfall damage pathways, maps exposure for critical infrastructure (power grids, arterial roads, medical shelters), and automates early-warning advisory dispatches for municipal and disaster management authorities in the Bay of Bengal / coastal APAC region.

The repo has two subsystems: the cyclone/infrastructure forecaster (the original Track 05 submission) and Sky Watch (a later addition, passive early-warning sky-scanning). Both are real, tested code — see Status below for exactly what has and hasn't been verified live.

Cyclone/infrastructure forecaster

Combines three real, independently-sourced signals into one Gemini-reasoned risk assessment:

1. Historical cyclone tracks — NOAA's IBTrACS North Indian Ocean basin dataset (real storm tracks, wind, and pressure records covering the Bay of Bengal, including India's own RSMC New Delhi agency data).
2. Live meteorological data — Open-Meteo's free forecast API (rainfall, wind speed) for the region being assessed.
3. Satellite imagery — Google Earth Engine, for the infrastructure exposure layer (what's actually in the storm's projected path).

Gemini (via the Interactions API) reasons over all three to produce a plain-language risk assessment and a draft early-warning advisory — Gemini does the judgment, the rest of the pipeline stays deterministic.

Sky Watch — passive early-warning scanning

Strictly civilian early-warning/alerting only — no tracking, targeting, or engagement capability of any kind, by design. alert_level and anomaly fields are informational flags for a human reviewer; nothing in this codebase acts on them automatically.

1. Day-sky vision (sky/vision_detect.py) — Gemini multimodal classification of a single camera frame for weather, pollution, and aerial-object anomalies (birds, aircraft, drones, debris, etc.). Classification only — no cross-frame tracking, no determination of intent.
2. Night-sky cross-referencing (sky/night_sky.py) — explicitly not vision-dependent: resolves real NASA GIBS VIIRS Day/Night Band tile URLs (VIIRS_NOAA21_DayNightBand) for a coordinate and date.
3. Hardware-agnostic camera ingestion (sky/camera_ingest.py) — ONVIF (the cross-vendor IP-camera standard) for discovery and stream resolution, plain RTSP as the universal fallback every camera brand supports. Hardware-unverified — see Status.
4. LLM routing across real data domains (sky/router.py) — Gemini classifies which of weather / cyclone-track / night-sky / research a query needs; plain Python then calls only those real modules. The model never fetches data itself.
5. RAG over public research (rag/) — arXiv's real public API (climate/disaster-response papers) embedded via gemini-embedding-001 into an in-memory vector store. Public, unclassified sources only.
6. Ontology/memory store (sky/ontology.py) — a flat subject-relation-object triple store, Gemini-extracted from observations/research text, JSON-file-persistable. Deliberately not a full RDF/OWL reasoner — see docs/ARCHITECTURE.md for why.

Status

Verified as of 2026-09-18. Read this section literally — it distinguishes "code exists and is tested against mocks," "verified against real external APIs," and "confirmed working live," which are three different claims.

Confirmed working live, right now:

• GET /health → 200 {"status":"ok"}, no auth required.
• GET /sky/night?lat=..&lon=.. → a real NASA GIBS tile URL, confirmed via a live request against the deployed service (not just local tests).
• GET /assess/{storm_name} → a real, successful, end-to-end Gemini-powered risk assessment. Confirmed live against AMPHAN/season 2020 at (13.08, 80.27): correct InfrastructureRiskAssessment JSON, reasoning that correctly referenced AMPHAN's real historical peak intensity (145 kt) and the actual live Open-Meteo forecast at the queried coordinates, and an honest confidence_caveats field noting zero satellite imagery scenes were available. This is the first real Gemini call to ever complete successfully in this build — the AI Studio billing block (see below) was resolved by using a separate, unbilled AI Studio project instead of the original paid/prepay-blocked one. Cold-start latency fixed: that first call took 2m41s, traced to the ~28MB IBTrACS CSV being re-downloaded from NOAA on every cold start (measured at 10-24s depending on network conditions). A snapshot is now bundled inside the installed package (verified present in the built wheel, not just the source tree) — local load time measured at 1.3s. Timeout fixed: a separate live test hung 90s+ with no error when Gemini returned a 429, because the SDK's timeout parameter defaults to unbounded. Every Gemini call (agent.py, vision_detect.py, router.py, ontology.py, rag/embeddings.py) now has an explicit 45s timeout, confirmed live: a rate-limited call now fails with a clear error in ~76s (one internal retry, each bounded) instead of hanging indefinitely. Known constraint, not a bug: the free-tier AI Studio project used to unblock billing has a hard cap of 20 requests/day for gemini-3.8-flash — confirmed via the API's own error message. Exhausted during this session's testing; resets on its own daily schedule.
• Earth Engine: real credentials authenticated locally, real Sentinel-2 queries confirmed against real coordinates through fetch_recent_imagery_count (locally — not yet wired into the live Cloud Run service, which is why the live /assess call above reported zero imagery scenes).
• IBTrACS and Open-Meteo: no auth needed, work as documented. One real bug fixed in this pass: IBTrACS uses a whitespace-only string as a missing-value marker in some cells, which crashed /assess with a 500 until fixed (see git history) — caught by this very live test, not by local tests alone.

Code exists, tests pass (47/47 locally as of the last full run), not yet exercised live:

• /sky/day-scan (day-sky vision classification)
• /research/index and /research (RAG indexing/retrieval)
• /route (LLM domain routing)
• These are Gemini-dependent but structurally identical in pattern to /assess, which now works live — high confidence they'll work once exercised, but "high confidence" is not "confirmed," so they stay in this section until actually tested live.

Explicitly not done:

• No physical storm-surge or rainfall-damage simulation — only data ingestion + Gemini's own reasoning, not a physical model.
• Camera ingestion is hardware-unverified — correct against the real ONVIF/RTSP library APIs (confirmed by introspecting the installed packages), never exercised against a physical camera, since none exists in this development environment.
• The RAG store is in-memory only (resets on every server restart) and needs an explicit /research/index call before /research returns anything.
• No demo video or pitch deck — non-code hackathon deliverables, not started.
• No AWS/GCP-style Partner registration or eligibility gate applies here (unlike this account's other hackathon repo, Omnidamus) — hack2skill has no equivalent partner-program requirement as of the rules checked.

Tools used

Purpose	Service	Package	Verified against
AI reasoning	Gemini API (Interactions)	google-genai	ai.google.dev/gemini-api/docs
Satellite imagery	Google Earth Engine	ee	developers.google.com/earth-engine
Historical cyclone tracks	NOAA IBTrACS (North Indian basin)	pandas (CSV)	ncei.noaa.gov IBTrACS v04r01
Live weather	Open-Meteo	httpx	open-meteo.com/en/docs
Night-sky imagery	NASA GIBS (VIIRS Day/Night Band)	httpx (tile URL only)	gibs.earthdata.nasa.gov, verified via live GetCapabilities and a live tile request
Camera ingestion	ONVIF / RTSP	onvif-zeep-async, wsdiscovery	introspected from installed package APIs
Text embeddings	Gemini (gemini-embedding-001)	google-genai	ai.google.dev/gemini-api/docs/embeddings + SDK introspection
Research corpus	arXiv public API	httpx	export.arxiv.org, verified live
Deployment	Google Cloud Run	gcloud	project project-16c62420-db89-448a-bab, region asia-south1

Setup

pip install -e ".[dev]"
cp .env.example .env  # fill in GEMINI_API_KEY and a GEE project ID
pytest

License

MIT — see LICENSE.