Selected work

Theo AI / Production system

Inbox to source-attributed case intelligence.

A durable AI pipeline for turning fragmented emails and documents into structured, traceable work product.

Role
Founding AI Engineer
Focus
Architecture + implementation
Domain
Legal AI
Timeline
2025 - Present
Email + documentsunstructured input
01Route
02Extract
03Normalize
04Classify
05Link
06Match
Case intelligencestructured + attributed

AI was not the difficult part. Keeping the evidence connected was.

Enterprise legal work arrives as threads, attachments, duplicates, inconsistent identifiers, and documents whose meaning depends on the matter around them. The system needed to recover that structure without losing the source trail operators use to verify an answer.

A lineage graph made intake traceable.

I designed and implemented the graph schema and persistence layers connecting messages, attachments, normalized documents, threads, and downstream case decisions. That graph became both an audit trail and a retrieval surface for matching new material to existing matters.

Normalize once

Content-addressed documents prevent repeated processing when the same file appears across a thread.

Narrow early

Deterministic identifiers and thread signals reduce the candidate set before model-based verification.

Fail closed

High-confidence paths attach directly; uncertainty falls back to the fuller matching workflow.

Keep provenance

Outputs preserve the path back to messages, documents, and evidence used to produce them.

Measured impact

52%lower benchmark processing cost
39%lower case-match cost
53%higher processing throughput

Measured on internal production benchmarks. Client names, corpus details, implementation thresholds, and proprietary prompts are intentionally omitted.

Next case study

Beehive