EPIWeather (역병예보) + AEGIRON — Epidemic Early Warning on Meta Model API

Infectious-disease early-warning system: a FastAPI inference engine (epiweather-api) plus a Next.js dashboard (epiweather-web). It fuses news, weather, mobility, genomic, and deforestation signals into outbreak alerts with honest provenance on every record.

Submitted to the Meta Global AI Developer Hackathon 2026 (Oct 26 – Nov 4, 2026).

• Track (proposed): Indie Developer
• Theme (proposed): Build for Your World — Meta's models in the public-health domain, solving an outbreak-detection problem firsthand.
• Demo video (3 min): link added before submission
• Live demo: run locally (see Quickstart) — no hosted demo claimed.

Built during the event (Oct 26 – Nov 4)

Per the event rule that the demo shows event-window work, the submission centers on what changed in the window:

1. Meta Model API extraction path (Oct 28, this branch's headline work)
   • epiweather-api/algorithms/meta_api.py (new): single source of truth for the Meta Model API convention — default muse-spark-1.3 on https://api.meta.ai/v1 (OpenAI-compatible chat/completions).
   • epiweather-api/algorithms/aegiron_model.py: extraction order is now NIM → Meta → local → Claude fallback; AEGIRON_META_FIRST=on forces Meta first for the demo.
   • Every extracted signal is tagged ai_provider=meta_model_api (+ data_sovereign flag); calls are observed in the capability ledger and surfaced on /api/status.
   • epiweather-api/tests/test_meta_api.py (new): 17 tests, all mocked (no outbound HTTP per repo policy).
2. WHO outbreak-declaration country evidence (#109, #111, Oct 26–27)
   • who_don_declarations.py + country_resolve.py: declaration announcements now carry resolved country evidence with logged basis.

How Meta's AI is used (beyond a basic integration)

• Muse Spark does production extraction, not a side demo: multilingual news text → structured outbreak records (disease, country, signal type) through the same prompt/schema as every other backend, so quality is comparable across providers.
• Honest provider tagging: each stored signal records which model made the label (extracted_signals.ai_provider), and /api/status separates "configured" from "recently observed working" — a claim this repo was built around after real incidents.
• Sovereignty gating: Meta-hosted inference is correctly marked non-sovereign; with AEGIRON_REQUIRE_LOCAL=on, sensitive data is blocked before leaving, and self-hosted OpenAI-compatible bases count as sovereign automatically.
• Unguessable-parameter discipline: the Meta path deliberately does not send NVIDIA-only parameters the team never validated against Meta.

Quickstart

Requirements: Python 3.11+, Node 20+, a Meta Model API key (dev.meta.ai → API keys).

# 1. API
cd epiweather-api
pip install -r requirements.txt
cp .env.example .env   # then set META_MODEL_API_KEY (and data-source keys)
uvicorn main:app --port 8000   # Swagger UI: http://localhost:8000/docs

# 2. Web (new terminal)
cd epiweather-web
npm install
cp .env.example .env.local   # NEXT_PUBLIC_API_URL=http://localhost:8000
npm run dev                  # http://localhost:3000

Key env vars (API .env):

Variable	Purpose
META_MODEL_API_KEY	Meta Model API key (MODEL_API_KEY also accepted)
META_MODEL_API_MODEL	Override (default muse-spark-1.3)
AEGIRON_META_FIRST	on = try Meta before NIM (demo forcing)
EPIWEATHER_API_KEY	Required: protects POST endpoints (fail-closed 503)

Verify the Meta path (1 minute, live)

cd epiweather-api
META_MODEL_API_KEY=<key> python -c "
import sys; sys.path.insert(0,'.')
from algorithms import aegiron_model as am
r, a = am.extract_signal_detailed('DRC 에볼라 신규발생', source='demo')
print(r.get('ai_provider'), '|', r.get('disease'), '|', r.get('country_iso3'))"
# expected: meta_model_api | 에볼라 | COD

Then open http://localhost:8000/api/status → capability_health → meta_text shows verified: true after a real call.

Layout

• epiweather-api/ — FastAPI service: main.py, collector.py, db.py, algorithms/ (inference layers), tests/ (2389 tests, mocked HTTP).
• epiweather-web/ — Next.js 16 dashboard (i18n, Leaflet maps).
• epiweather-handoff/ — engine sources reused by the API.
• docs/ — ops notes. DEMO.md — 3-minute demo script.

Test status

• New Meta path: 17/17 pass; affected area (AEGIRON/NIM/ledger/diagnostics): 220/220 pass.
• Full API suite: 2389 tests collected. test_main_endpoints.py needs localhost sockets, blocked in some sandboxes (fails identically without this change — environmental, not a regression).