Joshua Field
Defense has no codified standard. I built one.
GIDEON: AI case analysis for criminal defense, built by a 1,200-case public defender. Patent pending. CA Bar. Bay Area.
Despite the presumption of innocence, prosecutors have the playbook handed to them, forcing criminal defense attorneys to react. GIDEON changes that — intelligence honed by the professionals who use it, built on Atom-Structured Retrieval (ASR): it decomposes 241 California felonies into 126 atomic legal elements and scores every piece of evidence against every element the prosecution must prove. The attorney opens the file already knowing where to push.
The output isn't a chat window — it's defense work product: element-by-element evidence scoring, the contradiction map, the gap analysis. Sourced, and signed off by the attorney. The review step is load-bearing by design.
GIDEON gets more strategic with time: it maintains context within a case, across a caseload, and around the workflows and tactics of the attorney using it — because preparing a defender means knowing more than the law and the facts. It means knowing how to triage a crowded calendar and spend finite hours where they matter for the client in front of you.
Go-to-market: large public defender offices first — one procurement decision, hundreds of litigators, the highest fact-density caseloads in law. The private defense bar follows what the PD community validates.
GIDEON is my third system, not my first. The two underneath it were the R&D that made it possible — and hard to copy.
C3 — the context engine. Continuous Context Protocol: captures, distills, and re-serves everything you and the AI have done together, in any token budget. GIDEON's per-case, per-caseload, per-attorney memory runs on it. Provisional patent filed on the closed-universe baseline architecture.
UltrAI — multi-LLM orchestration. Routes prompts across Claude, GPT, and Gemini, runs multi-round cross-model refinement, and synthesizes a single response. Python/FastAPI. Patent pending.
Each new vertical inherits the trunk: same engine, new statutes.
Interested in learning more? Sign up here!
Drop your email and I'll keep you posted on releases, demos, and case studies.
I spent five years as a public defender — 1,200+ cases, 150 concurrent at peak (three times the recommended max), 80%+ success rate in trial-confirmed cases. The job teaches you one thing about the criminal justice system: defense has no codified standard. Prosecutors get the playbook; criminal defense attorneys are left to react. Then three election cycles running voter-protection and legal operations at national scale, and six years in legal policy. Running data-centric operations taught me the other thing: strong system architecture is what separates an operation that survives its own complexity from one that doesn't.
Before tech.Where I came from.
| Government | Obama Administration appointee (Peace Corps Press Director, secret clearance). Three election cycles. |
| Operations | Pennsylvania Voter Protection 2016 — 3,000 volunteers, 2,000+ Election Day incidents, zero legal challenges. Fair Census Project — $2.5M budget, 0.06% variance. |
| AI | MIT Certificate in Designing and Building AI Products and Services. Three production systems shipped. Provisional patents across the stack. |
Proof-backed
The architecture is grounded in a formal proof. 12 axioms, 5 theorems, 24 invariants, 19 failure modes accounted for. Most AI systems are "trust me." Mine is "check the proof." The proof does real work day-to-day: it keeps builds constrained to what the architecture actually permits, ties testing back to the user experience the system is supposed to deliver, and keeps maintenance targeted at the failure modes that actually matter.
Fractal
GIDEON is one instance of a more general engine. Point it at criminal defense and you get case analysis; point it at another knowledge-heavy field and you get that field's version. I built the core once and reuse it. Scaling it to a whole office, or to many firms, is the same engine running at a bigger scope.
Partnership-first
I don't build AI to replace people. I build the partnership layer between people and AI. The system gets sharper the more you work with it — learning your patterns and surfacing the right information and the right next move in the form you'll actually use. Every interaction is captured, distilled, and brought forward, so the AI you're using a year from now is the one shaped by everything you've done together.
The trust stack
In a privileged, bar-regulated domain, trust is the admission ticket — so it's architecture, not policy. Provenance: every finding traces to its source; every artifact carries an identity chain down to the file level. One chokepoint: a single typed egress point is the only path off the device — only atoms and typed placeholders cross, never narrative; per-firm key derivation; encryption at rest. Staged autonomy: the system earns autonomy through attorney corrections instead of promising it on day one. Audit: observable, logged, checkable. Don't trust me; check the proof.
Evidence
A particular method should produce particular results. Across nine patent experiments and ~960 controlled trials, the system architecture has been measured, not asserted.
- Distillates beat raw transcripts on retrieval quality with large effect (Cliff's d > 0.7).
- Friedman χ²(2) = 31.12, p < .001 across the three-way ranking comparison.
- Context Diffusion outperforms BM25 baselines on multi-hop queries with statistical significance.
- GIDEON ASR delivers 100% formula uniqueness across 126 primitives with 7.91-bit entropy.
- All experiments pre-registered; multiple-comparison corrections applied (Bonferroni–Holm).