PliegoAI：用 Gemini 评估西班牙公共采购机会的 B2B 产品

PliegoAI is a public, actively developed B2B product for the early assessment of Spanish public procurement opportunities.

It collects tender information, extracts requirements from source documents and compares them with a company's declared dossier. The result highlights potential eligibility risks, missing information and the evidence behind each conclusion so that a human can decide whether an opportunity deserves further review.

PliegoAI supports an initial assessment. It does not provide legal advice, guarantee eligibility or eliminate the possibility of model errors.

Product status

Public deployment: https://pliegoai.com
API health: https://api.pliegoai.com/health
Status: deployed and under active development
Adoption: personal testing only; no claim of established customers or commercial adoption
Product owner and developer: Luis Arias Mendoza

AI-testing-lab is the historical repository name. The product's final public name is PliegoAI.

What it does

Ingests public procurement opportunities and their documents.
Extracts structured requirements with Gemini.
Anchors model-produced citations to the source document.
Compares requirements with evidence declared in a company dossier.
Applies deterministic gates for objective blocking or review conditions.
Presents potential eligibility, uncertainty and evidence for human review.

Evidence and model controls

The application does not treat model output as automatically true. The current pipeline includes:

structured JSON output and strict Zod schema validation;
rejection of invalid JSON and unexpected output shapes;
exact citation checks against stored document text;
filtering of evidence references that do not exist in the supplied dossier;
deterministic eligibility gates outside the language model;
explicit UNKNOWN, CONFLICTING and expert-review states;
unit tests for fabricated citations, offset tampering and prompt injection.

These controls reduce unsupported output and make failures visible; they do not make the system infallible.

Technology

Frontend: React, TypeScript, Vite and Cloudflare Pages
API: Node.js, Express and TypeScript
Data: PostgreSQL, Drizzle ORM and row-level tenant isolation
Authentication: Supabase Auth with verified JWTs
AI pipeline: Gemini gemini-3.5-flash-lite
Delivery: Docker, Docker Compose and GitHub Actions
Testing: Node test runner, Vitest and Playwright

Repository structure

api/       Express API, persistence, extraction and qualification pipeline
frontend/  React application deployed through Cloudflare Pages
docs/      Architecture, product decisions and operational documentation
.github/   CI/CD workflows

Local development

API

cd api
npm ci
npm run typecheck
npm run test:unit
npm run dev

Copy the required variables into a local .env file. Never commit secrets or reuse production credentials for tests.

Frontend

cd frontend
npm ci
npm run lint
npm run test:unit
npm run dev

Integration tests require an isolated PostgreSQL test database and the test-only variables documented in the repository. Production credentials must not be used.

AI-first development approach

PliegoAI was built with an AI-first workflow. Luis defined the product requirements, architecture, validation rules, tests and deployment decisions, using Codex, ChatGPT and Antigravity to help implement and review the code. Generated code is treated as untrusted until it is understood, tested and validated against the repository rules.

See AGENTS.md for the engineering constraints and docs/ for the detailed technical documentation.

License

No open-source license has been granted. The source is publicly visible, but all rights remain reserved unless a license is added later.