Apprentice: small models that earn the job

Teams pay frontier prices for repeatable work: the same extraction, classification, or triage prompt running thousands of times a day. You wouldn't pay a master craftsman to make the same cut a thousand times a day. You'd train an apprentice. Apprentice (runapprentice.com) is that apprentice for your LLM stack. It watches the expensive model work, learns the task from verified examples, and takes over only when it proves itself. The loop: capture real input and output pairs, a human verifies rows into a gold set, GEPA optimizes the prompt, a small open model is fine-tuned on the verified rows, and an eval gate scores it on held-out gold rows. Only a model that beats the bar takes traffic. Fallbacks to the frontier count against the savings, never hidden. Rollback is one environment variable. The gate can say no, and that refusal is the product: cost cutting that never costs you quality. This is a real product, not a weekend demo. The console, Python SDK (pip install runapprentice), docs, an agent skill for Claude Code, Codex, and Copilot CLI, and a public benchmark with four tasks are all live. On real scanned receipts our fine-tuned 4B scored 89.17 against the GEPA-optimized frontier teacher's 79.58, on the same held-out rows, reproducible by anyone. For this hackathon we brought the loop to AMD: Gemma 4 E4B fine-tuned with a bf16 LoRA (TRL + PEFT) on an AMD MI300X with ROCm, served with vLLM on the same pod, Fireworks AI as the honest fallback. Measured July 10: raw 36.33, fine-tuned 61.67 on 60 held-out contract rows, +25.3 over the best published teacher score, about 8.4 GPU-minutes per adapter. The pod notebook with unedited outputs, the run report, and the trained adapter are all public. The market timing is real: OpenAI closes self-serve fine-tuning to new training jobs by January 2027, and those teams need somewhere to go. Trying Apprentice is free. We are taking 3 design partners for the migration.