Skip to content
Shonin
Beta

Decision Router Retrofit

We move the routing, scoring and approval calls in your agents off frontier models, and take a share of what you save.

The split

LLM writes
Keeps only the routes that generate text
System One decides
Takes over routing, classification, scoring and approvals
Code executes
Takes over counting, dates, limits and exact rules

Engine

We replay your logged traffic through the new decision layer, compare answers and cost, and switch only the routes that hold accuracy.

Rulebook

What correct means here, and the mistakes we've already caught. It grows one job at a time.

  1. R01Measure on your traffic, never on vendor benchmarks.
  2. R02A route moves only when agreement with the old answer clears the agreed bar on a held-out sample.
  3. R03Every moved route logs model version, probabilities and confidence.
  4. R04Thresholds are set per action, scaled to what being wrong costs.
  5. R05Anything that counts, sums or compares dates moves to code, not to a model.

Intake

  • Read access to agent code and a week of traces or logs
  • Current monthly model spend by route
  • The accuracy bar for each decision
UNIT
One agent or pipeline retrofitted
REVIEW LAYER
The client signs off on the replay report before any route switches.
DELIVERY
A pull request, a replay report with cost and agreement per route, and a monitoring view.
TURNAROUND
2 weeks

How we find buyers

  • Public teardowns: how much a real agent's bill dropped and why
  • A free savings estimate from one week of logs
  • AI builder communities and hackathons

More like this