AI Engineer
ADA Document Remediation Engine
Multi-critic AI document remediation engine in Python. Audit-logged QA loops, every model decision inspectable.
A production Python pipeline that ingests structured PDF input plus embedded media, classifies each document to a DocumentProfile, and runs text, table, and structure critics in parallel to evaluate the AI-generated remediation. An audit logger diffs between QA loops so every model decision is inspectable. The architecture ports to other image-and-data-to-generated-content pipelines.
The problem
AI-generated document remediation fails silently when the output is just text. Without parallel critics evaluating different aspects of the output, errors that are obvious to a human (a malformed table, a broken heading hierarchy, lost alt text) slip through. Single-pass remediation also makes it impossible to tell why a given model output was accepted or rejected when something goes wrong downstream.
What we built
The engine is a Python pipeline that classifies each input document to a DocumentProfile, then runs three critics in parallel (text, table, structure). Each critic evaluates the AI-generated remediation against its own criteria. An audit logger captures the diff between QA loops so every model decision is inspectable after the fact.
What was delivered
- Python remediation engine with DocumentProfile classification at the entry point
- Three parallel critics (text, table, structure), each with its own evaluation criteria
- Audit logger capturing diffs between QA loops so every model decision is inspectable
- Reusable architectural pattern for image-and-data-to-generated-content work
Outcomes
- Remediation errors caught before reaching the output stage
- Every model decision is inspectable through the audit log
- Parallel critic structure ports to adjacent generation pipelines without redesign
- Eval discipline that separates a demo from a system a team can ship
Services: Python, AI document remediation, Multi-critic eval, Audit logging, PDF + media pipeline