AIONTHEFLY Technologies logoAI FDE TOOLKITDeveloped by Dr. Sudin Baraokar

M07 · SLM · classifier · inference control plane

Intent routing and SLM

Send the cheap, certain work to a small model and reserve the frontier path for risk.

$ python -m fde_toolkit router --query "..."

Workflow

the order an FDE actually runs it in

  1. 1Classify intent, risk class and expected tool surface with a fast SLM.
  2. 2Score every candidate agent on accuracy minus cost, latency and risk.
  3. 3Abstain and escalate when confidence sits under the threshold.
  4. 4Dispatch to the specialist agent with a scoped tool manifest.
  5. 5Record routed cost against the naive frontier-model baseline.

Function output

deterministic trace

intent='current account balance' -> SLM  acc 0.97  lat 120ms  cost 1x  risk LOW  ROUTE
intent='investigate suspicious cross-border payments' -> AML multi-agent DAG  ESCALATE
batch of 8: routed $0.00412 vs naive frontier $0.03180 (-87%)