Reducto — Catalog Compliance Agent
Making high-volume catalog review an auditable decision flow
A seller catalog cannot rely on a language model simply saying “approved.” Every decision needs a defined shape, a confidence boundary and a clear path for ambiguous cases.
The challenge
The catalog receives more than 200,000 seller listing updates per day. Reviewing them means distinguishing routine compliant listings from cases that require a person to make the call.
Reducto is an asynchronous AI audit harness. It validates the structure of model decisions, auto-approves high-probability compliant cases and escalates ambiguous decisions to human auditors.
System approach
Separate the audit from the request
The workflow is asynchronous, allowing the listing update and its compliance evaluation to be treated as distinct work. That separation matters when individual evaluations take different amounts of time.
Constrain AI output
Strict schema validation makes the model return a decision that the surrounding system can interpret. A plausible paragraph is insufficient when downstream actions depend on a machine-readable result.
Keep a human escalation path
High-probability compliant cases can move through automatically, while edge cases are routed to auditors. The handoff is part of the product, not an exception to it.
Engineering considerations
These are design principles the work raises, rather than a claim about private architecture or unreported implementation choices.
Validate before acting
In any system built around probabilistic outputs, malformed, missing or unexpected fields should fail closed into a safe review state. This is a general design principle, rather than a claim about unpublished implementation details.
Model uncertainty explicitly
Treat “needs review” as a first-class outcome. A binary approve/reject interface can hide uncertainty and make operational auditing harder.
Measure the handoff
Useful measurements would include escalation volume, review latency and agreement between automated and human judgments. These are suggested evaluation dimensions; no internal values are published here.
Scope and outcome
The documented system handles a stream of 200,000+ daily listing updates with schema-validated AI decisions and a human review path. Approval rates, latency figures and financial impact are not public.
This is a public engineering narrative about an internal system. It does not reproduce production architecture, internal prompts, code or unpublished results.