We run the post-training loop for your AI agent workloads on infrastructure you control. Distillation cuts inference cost 10 to 20x. Reinforcement learning optimizes the small model on your task. Your data stays in your perimeter.
Every call to a frontier model sends your data to a third party, locks you into their pricing, and caps your margin. The model gets smarter every month but so does the bill.
More agents, more calls, more cost. There is no volume discount that beats running a 4B model on your own GPUs.
Every request, every response, every prompt flows through someone else's infrastructure. For regulated industries, that is a dealbreaker.
Switching providers means re-engineering prompts and pipelines. The model you depend on can change behavior overnight with no notice.
A drop-in replacement for your frontier API that gets cheaper and smarter every month, plus the machinery that makes that true.
A small library or proxy next to your existing agent. Logs requests and responses into a store you own. Open source. This is the open trace layer that feeds the commercial loop.
Pipeline code that distills teacher traces into a student model, then runs RL to optimize it on the task. Runs as jobs on your GPUs or sovereign cloud. Distillation first, RL second.
The eval harness scores the small model against the teacher on frozen test sets. Task metrics plus an LLM judge on a 0 to 5 scale. Produces a proof report: quality, cost per task, latency.
A serving container with an OpenAI-compatible API. Swap one URL. 80 to 90 percent of traffic goes to the small model. The hard 10 to 20 percent goes to a frontier fallback.
On top sits a simple dashboard.
Money saved this month. Quality score. Traffic split. One button: retrain with last month's traces.
That button is the subscription.
Four published blog posts tracking a 0.8B model through progressive distillation stages on a hard agentic benchmark. This is an open learning series, not a product benchmark. Each post breaks down the method, the data, and the honest results.
| Blog Post | Method | Score | |
|---|---|---|---|
| Part 1: Offline Teacher-Trace SFT | Hard-token SFT on teacher traces | 44 / 220 | Read → |
| Part 2: Offline Soft-Label KD | Top-k soft-label knowledge distillation | 55 / 220 | Read → |
| Part 3: Student-State Correction | Teacher correction SFT on student states | 67 / 220 | Read → |
| Part 4: On-Policy Probability Distillation | OPD with 35B teacher scoring student samples | 69 / 220 | Read → |
| Teacher baseline (GPT 5.5 medium) | Frontier model, no training | 115 / 220 |
A 0.8B student is 44x smaller than the frontier teacher. The scores show progressive improvement across training stages on a multi-turn SQL tool-use benchmark. Production deployments on narrower, well-defined tasks target 90 to 100 percent of teacher quality at 10 to 20x lower cost.
Free for the first 5 companies. Send 100 to 500 anonymized traces plus your monthly model bill. One week later, you get which steps a small model can take over, the projected cost cut, and the eval plan to prove it.
No commitment. No sales call. Just the report.
Not a pitch deck company. The full post-training loop has been published, step by step, with results. The open source trace layer is live and growing.
PhD in multi-agent reinforcement learning. Former CAIO at an acquired AI startup. Former Principal AI Scientist in agentic systems. 10+ years of enterprise ML delivery on AWS and Azure.
The Small Model Distillation blog series (Parts 1 to 4) is the startup's origin story: a 0.8B SQL agent trained through four distillation methods, published openly with benchmark results.