Mehdi Challakh

VLM: Gemini

LLM: Sonnet / Opus

OCR: Google Document AI

STT: Whisper

Backend: FastAPI (Python)

Jobs/queue: Redis + RQ

Auth: JWT + bcrypt

Frontend: React + Vite + TypeScript

Styling: Tailwind

Streaming: SSE / WebSocket

DOCX gen: python-docx

PDF export: LibreOffice headless

Containers: Docker + docker-compose

Doc editor: TipTap (ProseMirror)

File storage: disk volumes (MinIO/S3 later)

Embeddings: OpenAI text-embedding-3 (or BGE)

Database: PostgreSQL

Vector store: pgvector

Assistance for installers in creating and complying with CEE (Certificats d'Économies d'Énergie) files.

1/ VLM #1: Photo comparison for compliance

2/ Photo fraud detection

3/ VLM #2: Photo comparison for file completeness

4/ Compare quote information (scans) vs. database quote

5/ Detect/Extract/Verify equipment labels

6/ Detect/Extract/Verify stamp information

7/ Object detection and automatic labeling

8/ Fraud: Detect fake labels

A legaltech specializing in collective agreements and supplementary health insurance had an extraction pipeline that converted PDFs of contracts into bitmap images, sent them to an LLM, and then attempted to match the extracted benefits to a nomenclature of 286 entries. The result: frequent hallucinations, erratic matching between benefits and nomenclature, and unreproducible results from one run to the next on the same document. The audit identified the root causes at each stage. Bitmap extraction introduced variability from the start. Prompts combined multiple responsibilities in a single call. A filtering system by additional signals (keywords, embeddings) eliminated correct results before the LLM even saw them. And the nomenclature itself was unusable: 68% of entries lacked keywords, no descriptions, no structural relationships. The deliverables covered everything: documented architectural redesign (legacy vs. target diagrams), migration to structured JSON extraction via Document AI with block-level classification, audit and complete rewrite of prompts with an applied prompt engineering guide, comparative analysis of three additional signal architectures with a recommendation to remove filtering in favor of post-LLM verification, a plan to enrich the nomenclature (10 new fields, generation methodology), a context management strategy by specialty, and a 4-day sprint structured to establish a measurable baseline before any optimization. The team began implementing the recommendations upon delivery.