The Arbitrage Factory
Upload a schema, optimize it with DSPy/MIPROv2, deploy as a REST API — 10–30x cheaper than raw GPT-4o.
The Problem
Running AI in production at GPT-4o prices is expensive at scale. The insight most teams miss: GPT-4o-level quality is often achievable with a much cheaper model — if you optimize the prompts. DSPy's MIPROv2 does exactly that, but setting it up from scratch requires expertise in prompt optimization pipelines, synthetic data generation, evaluation harnesses, and deployment infrastructure.
The Build
Define your input/output schema in a UI. The platform generates synthetic training examples with GPT-4o, runs MIPROv2 optimization to find the best prompts for a cheaper student model (gpt-4o-mini), and produces a Regatta Report comparing cost and accuracy between baseline and optimized agents. The optimized agent deploys immediately as a REST endpoint with API key auth. Async job execution runs on Cloud Tasks with stuck-job recovery; compiled programs persist in GCS.
What Makes It Different
Most prompt optimization tools are research prototypes. The Arbitrage Factory is a production pipeline: schema builder, synthetic data generation, MIPROv2 optimization, cost/accuracy reporting, and REST deployment — all connected. The output is a live API, not a notebook. 146 tests passing.