Context
FDOT District 1 Letters of Response require detailed review for formatting, content, compliance indicators, and consistency across submissions. The review process can be repetitive and time intensive, especially when documents contain multiple page types, varying layouts, and requirements that involve both visual inspection and contextual judgment.
Challenge
The project required a system that could handle native PDFs, isolate relevant document sections, identify metadata, detect visual formatting issues, and summarize compliance findings without removing the human reviewer from the decision process. The workflow also needed to support secure web access and local execution.
My Role
I led the end-to-end AI workflow design and implementation. My work included PDF ingestion, page classification, title-block metadata extraction, LOR body isolation, automated margin and typography audits, visual overlay generation, LLM-based summarization, dashboard reporting, and user-facing delivery.
Approach
The system classifies and routes page types, extracts standardized metadata, isolates the main response body, and performs rule-based formatting audits. When violations are detected, the workflow generates visual overlays showing the location and nature of the issue. Structured JSON outputs are passed to an LLM for concise compliance summaries, then compiled into a program-level Excel dashboard.
Output
The project produced a secure review workflow with PDF overlays, structured JSON outputs, LLM-generated summaries, pass/fail indicators, flagged item tracking, an Excel reporting dashboard, a secure web portal, and a locally executable version for FDOT users.
Impact
The workflow turned a scattered manual review process into a structured AI-assisted QA/QC system. It helped reviewers validate issues faster, inspect visual evidence directly, and consolidate program-level results in one place.