AutoPilot AI

The AI assistant on this very site: a hybrid RAG pipeline where retrieval runs in your browser (MiniLM embeddings + BM25 + cross-encoder reranking via transformers.js) and generation runs on Tristan's own server.

Tristan fine-tuned Qwen3-0.6B with LoRA on Apple MLX (M3 Max), gated by a 40-question golden eval (FR/EN + refusal tests) before any checkpoint ships to production.

Inference: llama.cpp on a 2-vCPU Hetzner VM behind a FastAPI gateway (request queueing, rate limiting, SSE streaming, think-tag filtering), deployed with Docker + Dokploy + Traefik.

Graceful degradation by design: if the model is busy or unreachable, the chatbot silently falls back to extractive answers — the user never sees an error.