The problem
Bidding on French public contracts means wading through dense, rigidly structured consultation dossiers (RC, CCTP, CCAP, BPU, acte d'engagement) and producing a stack of strictly compliant response documents, while the relevant tenders are scattered across several official sources. The work is slow, repetitive and unforgiving of procedural mistakes.
What we built
- Dossier ingestion: upload the tender documents, then extract and analyse them (PDF, Word, Excel)
- A four-phase AI prompt pipeline, dossier analysis → administrative file → technical offer → financial offer, driven by a structured, versioned prompt library
- A model-agnostic engine orchestrating Google Gemini and Anthropic Claude with an OpenAI fallback, validating every output and tracking per-call cost
- A company profile (SIRET, qualifications, references) that feeds the responses; an earlier iteration aggregated tenders from the main official sources with alerts and a pipeline
Where it stands
The architecture is in place, a multi-provider AI engine, a prompt library covering the full bid workflow, document parsing, and an auth-gated, row-level-secured data model on a modern Next.js/Supabase stack. The product is at working-prototype stage, with the end-to-end generation loop being wired up.