Three craftsmen · Engineering, innovation, architecture
3MINDS is F. Andy Seidl, Alan Kucheck, and Bill French — a unique team with a seasoned, rich history in engineering, innovation, and technology architecture. Machines will sprint through a problem with perfect confidence and no stake in what breaks. That indifference is useful for speed. It is fatal for judgment. We keep the speed. We refuse the indifference. You hire three people who own the outcome — LEAP training that leaves your team literate in their own rules, or advanced AI systems that have to be right under load.
When the work demands it: LEAP, ontologies, and graphs. Tools — not the brand.
01 — The three
Complementary crafts, one stake: Andy’s ontology and systems architecture, Alan’s high-stakes engineering on platforms that cannot guess wrong, Bill’s agentic product architecture and how it lands in real organizations. Decades each shipping software that had to survive customers — different angles on why the work has to be right.
Serial software entrepreneur · Ann Arbor
Information architecture, search, and ontology systems — including the AIREX market stack. A tool-maker who builds the factories that build the products. Thirty-plus years designing and shipping commercial software; open-source systems craft (Dell OSS Mensa).
VP of Engineering · AI legacy-code archaeology
Leads engineering at Phase Change Software. Former air traffic controller — precision and conflict reconciliation under load. Decades bringing platforms like AIREX and AI-assisted legacy (COBOL/mainframe) work into production without theater.
Agentic architectures · St. George
Designs and deploys agentic AI architectures for enterprise outcomes — product, systems, and the written craft that makes judgment transferable. Phase Change Software; long trail of shipping integrations, SaaS, and AI systems people can actually operate.
A three-person engineering practice for high-stakes AI outcomes. Not a platform vendor. Not agent theater.
02 — The problem
Enterprises route deterministic work through unconstrained LLM loops. Latency, unit economics, and compliance all break. When the work needs systems, we separate domain truth, calibrated judgment, and deliberative synthesis.
03 — How we work
Cognition decoupled into deterministic domain truth, high-speed calibrated judgment, and deliberative synthesis — governed in a literate medium. The stack is how we build software that stays correct under load, scales without prompt-chain brittleness, and remains auditable years later — tools we reach for when stakes demand them, not the identity of 3MINDS.
Ontological Graph / TypeDB
Immutable business axioms, relations, and invariants live as formal relational logic and deductive inference — not semantic proximity. When rules are graph-native, every service shares one ground truth: a change lands in schema, not in a scatter of prompts and tribal memory.
That is how modern systems stay robust as the product grows. Correctness is typed, versioned, and enforceable — the foundation you scale from, not a wiki you hope people remember.
3MINDS Reflex Layer
Sub-100ms typed decisions — Choice, Score, Noul — on parameterized subgraphs with calibrated confidence. Judgment runs as a parallel forward pass instead of an autoregressive guess, so latency budgets and SLAs survive real traffic.
Confidence gates what executes automatically and what escalates. You scale decision volume without scaling hallucination risk — clever under load because the path is deterministic where it has to be.
Foundation LLMs
System 2 synthesis from foundation models, invoked only for generative or unstructured exceptions — never the inner state loop. Prose, redlines, and rare edge cases stay where language models excel; routing, invariants, and boolean checks stay out.
Clever generation without a hard boundary becomes unmaintainable debt. This boundary keeps the cleverness where it belongs, so the system remains inspectable when something goes wrong at 2 a.m.
LEAP
Human-in-the-loop literate specifications where domain experts and autonomous agents collaborate in version-controlled harmony. Rules live as readable markdown that compiles into ontology and reflex contracts — so legal, clinical, and engineering audit the same artifact.
Maintainability is not a later cleanup; it is the medium the system is written in. That is how software stays ownable years after the first deploy.
04 — Proof Points
Where sub-second, invariant-bound decisions are not optional.
In-line card fraud and transaction blocking under 100ms without breaking payment processor SLAs.
Automated Prior Authorization strictly bound to CMS guidelines and clinical coding ontologies.
Automated account renewal triage, churn defense, and expansion routing with zero human delay.
05 — Engagement
Phase 1 on-ramp: profile existing agent bottlenecks, map operational data onto systems that have to be right under load, and deliver a compiled LEAP specification ready for production sign-off.