Agent IA BH Assurance, Le Conseiller Augmenté

Le Conseiller Augmenté is an AI advisor for insurance agents, built solo for BH Assurance's month-long NEXT Challenge, where it won first place against teams of three. A front-desk agent reads each question and routes it to the right specialist: the product documents, a client's own records, or pricing. The hard part was never the model. It was the data.

BH Assurance's advisors lose hours digging through hundreds of product documents across six insurance lines to answer a single client. The source material is a mess: French, Arabic, and mixed, spread across inconsistent PDFs and spreadsheets. Off-the-shelf chatbots choke on it, because garbage in means garbage out.

1. 01

   Data first, model second

   Before writing any AI, I used a strong model to pull the entire product catalog out of dozens of messy, multilingual PDFs and spreadsheets and rebuild it as clean, consistent records. That groundwork is what made the whole thing work; without it, search just returns noise.

2. 02

   Retrieval that stays honest

   A question runs through semantic search, a re-ranking step, and careful context-building before the model answers, under a strict instruction to stick to what it retrieved and refuse to invent anything.

3. 03

   A front desk that routes

   A receptionist agent classifies each question and hands it to exactly one specialist, product knowledge, client data, or scenario analysis, so answers stay fast and focused instead of everything running every time.

4. 04

   The app around it

   A chat frontend and backend handle accounts and conversation history, one service looks up real client records, and another handles insurance quotes.

A Next.js frontend and an Express backend send every question to a receptionist agent, which routes it to one of four experts: retrieval over the product catalog, client-data lookups, scenario analysis, or pricing. Only the expert a question needs actually runs.

1. 01

   RAG's power comes from data quality, not model size. Investing a significant portion of the month into data augmentation, extracting from unstructured PDFs/CSVs, translating between French and Arabic, normalizing product categories, and structuring into rich JSON with keywords and key_points, was what made the system actually useful. The Gemini 2.5 Pro model was essential for this: the Flash model couldn't reliably parse the messy, multilingual source documents.

2. 02

   Going solo against teams of 3 forced absolute prioritization. I chose to build 4 focused services rather than one monolithic app, this let me test each service independently and iterate faster without the coordination overhead that teams face.

3. 03

   Qwen3-4B is a beast for its size via Ollama, the GGUF Q4_K_M quantization runs on consumer hardware while maintaining strong instruction-following. Combined with the 15K context window and strict factuality prompts, it answers insurance questions accurately without hallucinating.

4. 04

   FlashRank re-ranking was the biggest accuracy win after data quality. The difference between naive vector search and re-ranked results was dramatic, it eliminated the 'right category, wrong product' failure mode that plagued early iterations.

5. 05

   Docker Compose per service is the solo developer's secret weapon for hackathons. Each service has its own docker-compose.yml, which means I could demo any subset of the system even if other parts were still in development.