Anthony David

Internalized AI assistant project to simplify the work of business teams

• Use cases: simplifying language for user responses, supporting legal research, simplifying document writing.

• Complete RAG architecture: Ollama models locally and internal AI infrastructure in production, vector database for embeddings, SQL database for application data, LangChain (clients, chunkers, loaders, retriever), reranking by LLM, filters on metadata (collections, users, roles, session_id).

• Advanced chat: SSE streaming for real-time feedback, automatic use case detection, detection of business acronyms and retrieval of definitions, dynamic orchestration between chat and RAG mode with selection of relevant collections.

• Detection and classification of people, animals, and vehicles in video streams.

• Counting and tracking of objects in video streams.

• Super-resolution to improve the quality of distant images (SwinIR, ESRGAN, BSRGAN).

• DLStreamer pipelines with lightweight models for embedded systems.

• Grafana dashboards with MQTT traces for real-time analysis.

• Automated labeling via SAM2 in Label Studio (video tracking and object segmentation on images).

AI assistant composed of numerous tools and connectors, with a Graph RAG search engine

• Use cases: Human Resources, Purchasing, Legal, Documentation, code assistance.

• Implementation of tools: document counting, duration calculation.

• Development of connectors: Légifrance, Géorisques.

• Graph RAG improvements: graph database on a large set of contracts, reranking, query strategies according to use cases, Langfuse traces (LLM/embeddings/graph context), RAGAS metrics with Langfuse observability, prompt versioning.

• Model deployment: LiteLLM proxy for OVH AI endpoints (OpenAI format), deployment of an embedding model via vLLM on OVH VPS, Cohere rerank API compatibility proxy to vLLM API for a self-hosted reranker.