Parcourir les catégories
Explorer
Fiverr Pro
Français
$
USD
I will build a rag chatbot on your documents with sources and evaluation
Younes B
À propos de ce service
Most RAG demos fail in production because the retrieval is wrong, not the model. I build the retrieval right: semantic chunking, embeddings, hybrid dense+sparse search, re-ranking - and I measure it with an evaluation set, so every change is a number, not an impression.
What you get: a chatbot that answers from YOUR documents, cites its sources, and says "I don't know" when it should. Delivered as an API (FastAPI) with a web chat, Dockerized, with traces and cost per query.
Stack: Python, LangChain/LangGraph, Qdrant or Elasticsearch, OpenAI/Gemini/Mistral or open models on your infrastructure.
Recent work: production RAG engine for two retail brands (Mulliez group), OCR+RAG pipeline for an EventTech startup, document anti-fraud engine for a bank.
Senior AI Engineer, 6+ years, based in France. I lead teams of 2 to 5 engineers.
Découvrez Younes B
Younes B
Senior AI Engineer
- DeFrance
- Membre depuisavr. 2025
- Temps de réponse moy.1 heure
Langues
Français, Anglais, Arabe
Senior AI Engineer, 6+ years, focused on generative AI that runs in production: RAG pipelines (chunking, embeddings, hybrid search, re-ranking), multi-agent workflows (LangGraph), LLM-vision OCR and document extraction, plus the observability that goes with it: traces, continuous evaluation, cost/latency trade-offs. Full-stack and cloud end to end: Python/FastAPI, TypeScript/Next.js, GCP, Azure, Kubernetes, Terraform. Recent work: OCR+RAG pipeline for an EventTech startup, production AI engine for two retail brands, document anti-fraud engine for a bank. I lead teams of 2 to 5.
Mon portfolio
FAQ
Which LLM do you use?
Your choice: OpenAI, Gemini, Mistral, or an open model hosted on your side for confidentiality. I recommend one based on your documents and budget.
My documents are scanned PDFs. Is that a problem?
No. LLM-vision OCR is part of the pipeline: scans, photos and non-standard layouts are handled.
Where does the chatbot run?
In Docker on your server or your cloud (GCP, Azure, AWS). Nothing stays on my side after delivery.
How do I know it works?
You get an evaluation set: your questions, the expected answers, and a score per version. Every change is measured, not guessed.
