Contractor AI compares every public contract it ingests against thousands of similar historical contracts and flags the ones whose price doesn't add up — using open data, transparent statistics, and no black-box "trust me" scoring. It currently covers eight countries: Paraguay, Colombia, Costa Rica, República Dominicana, Perú, El Salvador, Brasil, and Uruguay (18,000+ contracts and counting), with a public dashboard, an institution-ranking board, a channel for citizens to flag or add context on any contract, and a live tool to check whether a contract that isn't in the dataset yet looks priced fairly.
It's built for the OEA's Desafío de Datos para la Democracia (25 years of the Inter-American Democratic Charter) and designed to keep working — and scaling to more countries — long after that.
Live app → · API docs → · Run it locally ↓
| 🔎 Anomaly detection | Every contract is scored with a modified z-score + Tukey IQR fences over log(amount), computed against peers grouped by buyer → category → country (whichever has enough data). Robust to outliers, fully explainable — no ML confidence score pretending to be more certain than it is. |
| 📄 Analyze a new contract | Upload a PDF, paste a link, or (soon) a photo. Text/amount extraction runs server-side; the amount gets compared live against the same statistical baseline the rest of the app uses. Nothing here claims to run a live AI model — see backend/app/analysis.py's docstring for exactly what is and isn't happening. |
| 📊 Dashboard | Aggregate metrics, contracts/anomaly-rate trend by year, top categories, a cross-country comparison, and CSV export. |
| 🏆 Institution ranking | The procuring institutions with the strongest track record (lowest anomaly rate, minimum sample size) — a "who's doing this well" board, not a shame list. |
| 🗣️ Citizen participation | Anyone can flag a concern or add context on any contract, no login required. |
| 🌙 Light/dark theme | Light by default, with a persisted toggle. |
frontend/ Next.js 15 (App Router) + React 19, plain CSS with custom
properties for theming — no Tailwind/shadcn dependency.
backend/ FastAPI + SQLAlchemy 2.0, Postgres in production / SQLite
for local dev. Read-only endpoints, a live comparison
endpoint, a dashboard aggregation module, citizen reports.
backend/scripts/ One ingestion script per country's data source (live
APIs where available), plus the statistical anomaly job that
scores every contract that doesn't have a signal yet.
docs/ Architecture plan, ADRs, and a live PROGRESS.md/CHANGELOG.md
pair documenting what's actually been built and why.
Full design rationale: docs/architecture/PLANNING.md ·
decision records: docs/adr/ · what's built and what's next:
PROGRESS.md / CHANGELOG.md.
| Country | Source | Notes |
|---|---|---|
| 🇵🇾 Paraguay | DNCP historical dataset (OCDS) | Amounts CPI-adjusted to a reference year for the original prediction model — see backend/README.md. |
| 🇨🇴 Colombia | SECOP II live API (datos.gov.co) + a processed third-party dataset | No verified per-contract exchange rate — shown in original COP. |
| 🇨🇷 Costa Rica | SICOP / Observatorio de Compra Pública, live | Updated daily; USD conversion done by the official source. |
| 🇩🇴 República Dominicana | DGCP live API | Shown in original DOP. |
| 🇵🇪 Perú | OECE live API (gob.pe/oece) | OCDS format; category distribution supports /analyze/compare by category (≥3 groups, <60% concentration). |
| 🇸🇻 El Salvador | COMPRASAL live API (DINAC) | Dolarizado desde 2001 (USD nativo); row-level data (line items, not full contracts); personal data excluded per privacy policy. |
| 🇧🇷 Brasil | Compras.gov.br open-data API (dadosabertos.compras.gov.br), live |
Federal contracts. The API requires a codigoOrgao per query — there is no whole-dataset endpoint — and caps date windows at 365 days, so ingestion walks organisations one at a time. Shown in original BRL. |
| 🇺🇾 Uruguay | ARCE via the national open-data catalogue, native OCDS 1.1 | Yearly OCDS archives from 2002 onward. Amounts live on award line items (quantity × unit.value.amount), not on the award, and the data is multi-currency (UYU, USD, EUR, UYI); USD awards keep a real amount_usd, the rest stay in their original currency. |
Every anomaly is scored two independent ways where data allows: an NLP model's deviation from its predicted price (only where a precomputed prediction exists), and an always-available statistical method (median + MAD) that runs on every contract regardless. See ADR 0003 for why those two signals are kept deliberately separate instead of blended into one score.
Requires Python 3.12+ and Node 20+.
# backend
cd backend
python -m venv .venv && source .venv/Scripts/activate # .venv/bin/activate on Linux/Mac
pip install -r requirements.txt
python scripts/migrate_dataset.py # loads Paraguay's historical dataset into SQLite
uvicorn app.main:app --reload --port 8000
# frontend (separate terminal)
cd frontend
cp .env.local.example .env.local # NEXT_PUBLIC_API_URL=http://localhost:8000
npm install
npm run devOpen http://localhost:3000. API docs at http://localhost:8000/docs.
Full details, environment variables, and how to pull in the other 3
countries' live data: backend/README.md ·
frontend/README.md.
cd backend
pip install -r requirements-dev.txt
pytestRuns against an isolated, throwaway SQLite database — never against
whatever DATABASE_URL your .env happens to point at, even if that's a
real Postgres instance. See backend/tests/conftest.py.
Public write endpoints (link-based contract extraction, citizen reports)
get real scrutiny: an SSRF guard with redirect re-validation, rate
limiting, honeypots, and pinned dependencies. Full writeup, including the
one known residual gap that's disclosed rather than hidden:
SECURITY.md.
Contractor AI started as a Streamlit prototype fine-tuning a multilingual BERT model to predict fair pricing for Paraguayan public tenders, aimed at helping MSMEs bid more competitively — built with VIGIA, ASEPY, and REACCIÓN in Paraguay, informed by prior work at Observatorio Fiscal (Colombia). That prototype is still live at hack-corruption-contractor.streamlit.app. It's since grown into the public, multi-country platform described above — see ADR 0001 for why, and how the two coexist during the transition.
CEO & Forward Deploy Engineer — Yefry Nunez · CTO & AI Engineer — Natalia Ramírez Pérez · Chief of Information — Domingo Aybar Santos · Chief of Politics — Nicole Checo · Chief of Financial — Jomayris Rosario Medina · plus Cristian Sosa, Dayanni Olivo, Daniel Duque, and Daniel Sosa. Full profiles on the live site.