↓ Skip to main content
Background Image

DistillAudit

·123 words·1 min
Inference-Engineering Startup Distillation Audit Governance
Manoj
Author
Manoj
ML Engineer @ 7-Eleven
Table of Contents

Pitch
#

DistillAudit checks whether a distilled model inherited unwanted preferences, biases, or behavioral quirks from its teacher.

The product is a test suite, report, and ongoing monitoring service for companies releasing smaller models.

Workflow
#

  1. Connect teacher and student models.
  2. Run structured probes across sensitive behavior categories.
  3. Compare response distributions.
  4. Flag unexpected transfer not present in the explicit training data.
  5. Produce a model governance report.

Why Now
#

Distillation is becoming routine because inference cost matters. Governance teams need a way to say not only “the student is accurate” but also “the student did not inherit hidden behavior we cannot explain.”

Risks
#

  • Compliance sales cycles are slow.
  • Probe coverage is always incomplete.
  • The product needs credibility, benchmarks, and probably academic validation.