AI Document Processing Platform

A senior analyst opens a 200-page transfer pricing report on Monday morning. By Wednesday afternoon, she has extracted forty data points, cross-referenced three appendices she almost missed, and flagged two contradicting figures on pages 47 and 183. That workflow -- plan what to extract, retrieve evidence, flip back to verify, write it up -- became the blueprint for every pipeline I built.

The problem was never "can AI read a PDF." It was that analysis requires judgment: knowing when evidence is sufficient, when to switch search strategies, and when to stop trusting text and look at the chart. I designed a dual-mode adaptive RAG system where an AUTO router selects PASSIVE single-pass retrieval for simple queries or AGENTIC self-reflective loops for dense analysis. Users never choose a mode.

The extraction v2 pipeline follows six stages: plan, retrieve, evaluate-and-flip, generate, verify, finalize. After initial retrieval, the evaluate-and-flip node examines coverage gaps against the extraction schema and switches between dense semantic search and sparse keyword retrieval depending on what is missing. A concept not found semantically might surface through an exact term buried in a footnote.

The Q&A pipeline adds a reflect node that critiques answers against source passages, catching hallucinations before they reach the user. The comparison pipeline normalizes terminology across documents before diffing -- essential when two reports name the same entity differently. A ConfigurableAIClient with database-driven provider selection lets customers control which APIs touch their data, swappable via the admin UI without code changes.

Financial documents contain charts, scanned attachments, and tabular images that text parsers silently skip. I integrated VLM invocation on demand -- the pipeline routes visual content through a vision model only when needed. For aged scans with faded ink, I apply CLAHE preprocessing before passing images to the VLM, improving accuracy on degraded inputs without hurting clean ones.

• What surprised me: the evaluate-and-flip node had more impact on extraction quality than any prompt engineering. The right retrieval strategy per field mattered more than the right words to ask the LLM.
• What I would do differently: build the comparison pipeline's terminology normalization as a shared service from the start, rather than duplicating alignment logic before refactoring.
• What stuck with me: enterprise AI is not about model capability. It is about giving analysts transparent evidence trails so they trust the output enough to act on it.