Document Integrity & Traceability
Human-in-the-Loop Anomaly Detection & Cross-Document Verification
Identifies provenance gaps, competing versions, metadata conflicts, and version drift across high-consequence records without converting uncertainty into unsupported certainty.
What was difficult, fragmented, or broken?
In contractual negotiations, loan underwriting, regulatory filings, and complex litigation, multiple parties exchange dozens of iterative drafts and exhibits. Minor edits, backdated signatures, deleted clauses, and metadata conflicts slip past human reviewers. When AI systems are deployed to summarize changes, they frequently hallucinate consensus or smooth over critical contradictions.
Why the obvious solution wasn't enough
Conventional "track changes" and PDF diff utilities only highlight character-level additions and deletions. They cannot detect semantic discrepancies (e.g., conflicting loan payoff figures across three schedules) or timeline impossibilities (e.g., a notarization timestamp preceding document creation). Generative AI summaries tend to synthesize a coherent narrative by silently dropping contradictory details.
My Role
What I Built
- A cross-document contradiction scanner that extracts structured entities (dates, monetary amounts, property addresses, party names) and flags discrepancies across files.
- A provenance and drift ledger that maps the evolution of key clauses across multiple iterations.
- A human-in-the-loop review interface that presents highlighted conflicts alongside original source snippets for rapid verification.
- Deterministic validation rules that preserve uncertainty—explicitly surfacing unresolved ambiguities rather than forcing an AI-generated guess.
Draft Iteration A ────┐
Draft Iteration B ────┼──► [Schema Normalization] ──► [Entity Cross-Check] ──► [Conflict Flagged] ──► [Human Approval]
Notarized Exhibit ────┘ │
Preserves Contradiction / EscalatesHow I Approached It
My core architectural principle is: AI should help identify what deserves human review without converting uncertainty into unsupported certainty. I designed the pipeline to parse unstructured documents into standardized JSON schemas, calculate semantic and numerical variances, and route any deviation above zero tolerance directly to human verification.
AI is an instrument, not an authority.
Used multi-model evaluation across leading frontier reasoning models to benchmark interpretive consistency when auditing filings. When models disagreed on whether a clause had changed materially, that exact divergence was treated as a signal that the clause was ambiguous and required human inspection.
Key Decisions & Tradeoffs
Explicitly prohibited automated document resolution: the system highlights and frames conflicts, but only authorized human reviewers can approve an authoritative state.
Engineered granular diffing at the semantic clause level rather than word-level, filtering out harmless formatting noise (whitespace, line wraps) to highlight actual risk.
The Result
Developed a defensible verification system that flags hidden document anomalies, implemented across mortgage file audits and legal exhibit preparations.