Streaming RAG platform for logistics

At Adilas Tech I lead AI development on a logistics platform whose core is an end-to-end streaming RAG application. Operators ask questions in plain language and get grounded answers drawn from the platform's own data, streamed into the interface as they generate.

I architected the full path: a React frontend connected over SignalR to a high-throughput FastAPI inference service running vLLM, with asynchronous retrieval against a Qdrant vector store. The streaming protocol was designed so the frontend renders tokens the moment they exist, keeping perceived latency low even on long answers.

The platform runs a hybrid model strategy that I validated and implemented: a routing layer decides per query whether a self-hosted model can answer or whether the request should escalate to GPT-4o. That keeps costs predictable without capping quality on hard queries.