Knowledge + MLOps
DomainForge
Fine-tune behavior, retrieve facts — governed eval gates on both.
End-to-end support triage: capstone SOP RAG, Bitext QLoRA + DPO, S0→S4 eval ladder, Ollama bench UI, adapter registry — RAG holds facts, PEFT holds JSON discipline.
In one sentence
Governed support triage — RAG corpus + QLoRA JSON envelope + S0–S2 eval compare.
Decision
RAG for facts, LoRA for schema — never memorize policies in weights.
Measured signal
S0→S4 compare · DPO preference win-rate · `/bench` Ollama metrics UI · `VLLM_BASE_URL` Path B.
Honest limitation
Mistral QLoRA requires GPU; Render API uses mock/template baseline unless Ollama/vLLM Lab wired. Path B is educational (not CUDA multi-LoRA).
- 13 SOP docs — hybrid governed retrieval
- Bitext → QLoRA SFT + DPO alignment (S3/S4)
- S0→S4 compare harness + preference pairs
- Ollama bench UI · GPU pipeline docs · educational vLLM Path B (ADR-022)
Architecture diagram
FastAPI · TRL · PEFT · Chroma · Next.js · Vercel · Render