TenderAI

Inspiration

50% of Moroccan SMEs are eliminated from public tenders not because they lack the skills, but because of missing or incorrectly formatted documents. A qualified company losing a contract to a paperwork error is not just frustrating — it's an economic problem. We wanted to fix that.

What it does

TenderAI transforms a public tender PDF (appel d'offres) into a fully compliant response dossier in minutes.

The user uploads the tender document and enters their company profile. The platform automatically:

• Extracts all technical, administrative, and financial requirements
• Generates a complete response in administrative French
• Validates the dossier against Moroccan procurement law (Décret n° 2-22-431)
• Scores compliance from 0 to 100
• Flags disqualification clauses in red before submission
• Auto-refines the dossier if the score falls below 60

How we built it

• FastAPI — async backend, handles the full AI pipeline
• PyMuPDF — extracts clean text from complex French administrative PDFs
• Gemini 1.5 Pro — reads the full tender document and generates the response dossier
• Featherless (Qwen2.5-32B) — validates the generated dossier and produces the compliance score
• React + Vite — three-screen frontend: Upload, Loading, Results

The pipeline runs as a chain of specialized AI agents: extraction → generation → validation → refinement loop if needed.

Challenges we ran into

• PDF complexity — Moroccan administrative PDFs are dense, inconsistently formatted, and sometimes scanned. Getting clean text extraction required significant tuning.
• Administrative French tone — Generic LLM output sounds nothing like real Moroccan procurement language. Prompting the right formal register took multiple iterations.
• JSON stability — Getting the validation model to return clean, parseable JSON every single time at temperature 0.1 was harder than expected.
• Regulatory accuracy — The Décret 2-12-349 was abrogated in September 2023 and replaced by Décret 2-22-431. Most existing tools and datasets still reference the old law. We had to manually verify and hardcode the correct elimination rules.

Accomplishments that we're proud of

• A working end-to-end pipeline that goes from raw PDF to scored dossier in under 2 minutes
• A compliance score that actually reflects Moroccan procurement law, not generic evaluation criteria
• An auto-refinement loop that improves a failing dossier without human intervention
• A demo that works live with real appels d'offres from marchespublics.gov.ma

What we learned

• Prompt engineering for legal and administrative content is a discipline of its own — precision matters more than creativity
• Multi-agent pipelines fail at the weakest link. Stability at every step matters more than peak performance at one step
• The real problem in public procurement is not writing the offer — it's knowing what will get you eliminated before you even start

What's next for TenderAI

• Tender monitoring — automatically track new appels d'offres matching the company's sector
• Document checklist generator — tell the user exactly which physical documents to prepare before submission day
• Success analytics — track which tender types the company wins and why
• Multi-user workspace — allow the full proposal team to collaborate on a single dossier
• Arabic language support — extend to Arabic-language tenders published by Moroccan municipalities

Built With

• docker
• fastapi
• featherless
• gemini
• python