arrow_backBack to What I Build

Legal AI

Your research tool skips what it never indexed.

Most legal AI is a search box: fast, confident, and capable of inventing citations. I build legal research systems where retrieval is deterministic and every claim is anchored to a real statute or opinion. It found 100% of the relevant law on the bar exam, and beat the $1,200-a-seat incumbent on the exam itself.

What Usually Looks Normal Until I Point At It

Cloud legal AI hallucinates cases and nobody catches it until a filing

Statute recall below 100% is treated as "good enough" for client work

Privileged client communications route through third-party LLM servers

Incumbent tools cost $1,200 a seat and still miss relevant law

How I Build It

01

Classify first, then retrieve — a 4-dimensional router maps the situation to exact statute sections

02

Citation-aware retrieval with small-to-big chunking and co-citation graphs

03

Local or self-hosted inference so privileged data never leaves your perimeter

04

Adversarial benchmark: every build is tested against the hard cases

Outcomes

100% of the relevant law found on the bar exam — the $1,200-a-seat incumbent lost the head-to-head.

Runs on hardware the firm owns for a few dollars a month

Technology Stack

PythonFastAPISQLite+FTS5PostgreSQLLocal LLMsCitation graph

Capabilities

Dimensional Classification Router

Instead of just searching for keywords, the system understands the who, what, where, and why of your case and points you straight to the laws that apply. It found 100% of the relevant law on the bar exam, and beat the $1,200-a-seat incumbent on the exam itself.

How It Works

  • Parse the fact pattern into 4 classification dimensions
  • Map each dimension combo to statute sections via lookup tables and override rules

Outcomes

  • 100% of the relevant law found on the bar exam — the $1,200-a-seat incumbent lost the head-to-head.
  • Instant answers running on a normal computer

Citation-Aware Retrieval

I index opinions, verdicts, and statute links into a citation graph so retrieval surfaces authorities that actually cite the relevant law — not semantically similar but legally irrelevant documents.

How It Works

  • A corpus of 1,002,364 court opinions with 18.1 million citation links and statute links
  • Extract 14,889 real jury verdicts via deterministic regex pipeline

Outcomes

  • No hallucinated cases
  • Comparable verdict data for case valuation

Judge Analytics

It knows your judge: their plaintiff-versus-defense win rate, the statutes they cite most, the damages they have actually awarded, and who appointed them. Profiles built from an AI scan across 67,028 indexed opinions — not a directory listing.

How It Works

  • Political and appointment background linked to ruling history
  • Win/loss, affirm/reverse, practice-area breakdown, most-cited statutes, and extracted damage awards per judge

Outcomes

  • A judge profile beside the research, not a separate subscription
  • Motion strategy informed by how this judge actually rules

Private Inference

Legal data does not need to leave your network. I deploy classification and retrieval on local or dedicated inference so privileged communications stay in your perimeter.

How It Works

  • Private classification on hardware you control
  • Self-hosted vector and full-text indices

Outcomes

  • CISO-clean data handling
  • Runtime cost drops to ~$1/month

The research is public.

Benchmarks run on machines I own, measured against frontier systems — with the failed approaches reported next to the results. Read it before you take my word for anything.