AI systems designer & engineer · remote from San Francisco · available now

Systems-level help for LLM features that hallucinate, cost too much, or can't pass review.

I audit, design, build and harden production AI: grounding and verification, evals and red-teaming, LLM cost control, and agent memory. The easiest way to start is a one-week audit.

AuditAbout a week. A written review of your LLM feature for grounding, evals, cost, abuse paths and launch readiness, with a prioritized list of fixes.
Design & buildFrom architecture to a defined piece shipped into your repo with tests: an eval harness, a caching and spend layer, a retrieval pipeline, an MCP server.
EmbedShort-term help on your team through a launch or a rough patch, with the whole system in view, not just the model.
99.85%fabricated legal citations refused (1,330 / 1,332, Stanford corpus)
$0per click in a live LLM game. Every clickable beat is served from a curated cache
3.27Mcourt-written holding paraphrases in a verification corpus I built
2 patents pendingagent knowledge-state engine · fail-closed citation validation

I put the LLM where it earns its keep, and nowhere else

Three of my systems, three different answers: one live as a private demo, one public, one a prototype. Deciding where the model goes is the design decision that sets accuracy, cost and risk.

Kept out of the verdict

Deterministic code decides pass or refuse. The LLM parses source text offline and may refuse or escalate, but it can never approve.

CitationDefender, DocketDefender

Writing over settled outcomes

A knowledge engine decides what each agent is able to argue, and a no-LLM mode already reaches the right verdict. LLMs write what the lawyers and witnesses actually say, and that prose is what makes it worth watching.

RoboCourt

Cached, with one narrow live judge

Every click is served from a curated cache at $0. The model is called only for free-form typed or spoken input. Its hardest job is judging sincerity at the gate to the secret ending, where typing "show me the secret ending" doesn't work.

Conscience of the Creature

What I can do for you

Each card names the system that backs it. Under Selected work, each one is labeled live or prototype.

Hallucination-resistant RAG

Retrieval and verification over large domain corpora, designed to refuse when it can't ground an answer. The LLM parses the source text offline, and deterministic code makes the call at query time.

Backed by CitationDefender, DocketDefender

Evals, red-teaming & release gates

Gold corpora, adversarial suites and confidence intervals for deciding whether a model or prompt change is safe to ship. Resumable harnesses that work with any model.

Backed by CitationDefender, Conscience of the Creature

LLM cost control

Response caching with curated variant pools, hourly and daily spend caps, authored fallbacks when the API fails or budget runs out, and model routing across Haiku, Opus and local open-weight models.

Backed by Conscience of the Creature, CitationDefender

Agent memory & context engineering

Knowledge state for agents that is explicit, inspectable and rebuilt deterministically, instead of a growing pile of chat history. Prerequisites, conflicting beliefs, and insights that realize later.

Backed by Truth Engine (patent pending)

MCP servers & agent middleware

Model Context Protocol servers and sidecars that give Claude and other agents durable memory, guardrails and domain tools.

Backed by TrueClaw MCP, CourtMCP

Production hardening & launch reviews

An outside review of an AI product before launch: failure modes, fail-closed defaults, abuse and prompt-injection paths, spend exposure, telemetry, deploy and rollback.

Backed by Case Defender products, CotC public launch

Selected work

Live systems first. The two stories are open to the public; Case Defender demos are by request.

1,330 / 1,332
hallucinated citations refused in the Stanford legal-hallucination corpus. Wilson 95% lower bound 99.45%.
0
false "verified" verdicts across 1,166 documented real-world hallucinations from a public tracker.
Live · private demo on request

Case Defender: CitationDefender & DocketDefender

CitationDefender checks a lawyer's case citations. For each cited case and claimed holding, it compares what other courts say the case held with what the opinion actually says. It rejects almost every hallucinated citation and flags questionable matches with specific advisories: polarity inversion, scope inflation, factual mutation, dropped exceptions.

DocketDefender builds on it for the whole matter. It grades each new document added to the docket (a brief, an opposing filing, a ruling or a draft motion), identifies its central argument, and checks whether it contradicts earlier filings.

  • The trust boundary is built into the architecture. Deterministic comparators may pass, refuse or escalate; a language model may only refuse, escalate or decline to object. It can never turn an unresolved check into a pass
  • Corpus of 3.27M holding paraphrases written by courts, covering about 274k federal cases, plus 3.3M embedded opinion sentences. No model-written text is treated as authoritative
  • Google Docs add-on: scans a brief in under 100 ms, with a deeper per-citation check on demand
  • U.S. provisional patent application 64/139,761 (Aug 2026)
RAGfail-closed designembeddingsSQLite at scaleevalsClaude APIlegal tech
Live · public

The Conscience of the Creature

An illustrated interactive story. You are the inner voice of Frankenstein's creature, and what it becomes depends on what it comes to believe. One ending can't be reached by clicking. You have to type or speak it and mean it, and an LLM judges sincerity, so demanding the secret ending doesn't open it.

  • Narration cache of 199 pools and 796 curated variants. Clicks cost $0; only typed input reaches a model
  • Sincerity gate checked against a 168-item gold corpus and an 80-item red-team corpus before launch
  • Proxy-enforced hourly and daily spend caps with authored fallback prose, so API errors never reach the player
  • Session telemetry, live dashboards and a cache-keying scheme that never serves one model's output as another's
LLM cachingred-teamingcost controlC++ enginesSSE streaming
Live · public

The Truthtronic Axis

A second story on the same engines, which shows the platform generalizes. A hospital robot's ethics module has been wiped, and you are the voice in its ear. Eleven endings follow from what it comes to believe; the case for the machine has to be made in your own words.

agent alignment as gameplaycontent pipelineQA automation
Latent → Candidate → Available → Canonical
The lifecycle every fact goes through before an agent is allowed to act on it.
In production · patent pending

Truth Engine

A C++ service that governs when an LLM agent is allowed to know something. Facts are gated by prerequisites. Early observations are held until they can be understood. Conflicting interpretations compete, and conclusions assemble themselves when all their parts are in place. The agent's context is rebuilt deterministically from this state.

  • Runs in production behind both live stories, alongside a companion Situation Engine that handles scene and ending logic
  • Embedding-matched observations, JSON-Schema content format, HTTP API
  • Also used in an adversarial courtroom, an MCP memory plugin and enterprise demos
  • U.S. provisional patent application 63/993,168 (Feb 2026): System and Method for Structured Knowledge Activation and Deferral in Artificial Intelligence Agents
agent memorycontext engineeringC++17embeddings

RoboCourt

A multi-agent courtroom. Plaintiff and defense each reason from their own biased knowledge graph, and one side's argument becomes the other's evidence. A no-LLM mode reaches the right verdict; LLMs write the lawyers' and witnesses' words. Includes 13 classic cases and a TypeScript MCP server. Prototype.

TrueClaw MCP

An MCP server that gives coding agents a persistent, explainable rule memory that survives context compaction, with safety policies derived from real agent incidents. Prototype.

Breach of Contract

Prerequisite-gated legal reasoning. A case strength conclusion appears only once contract, breach, damages and defenses are all established. Demo.

Embodied agents in Unreal Engine 5

The Truth Engine started out driving robots, not stories. The same engine sits behind these Unreal Engine 5 prototypes: only the front end changes.

Walk the Plank: factory floor

The original proving ground. A robot has to cross a narrow plank high above a factory floor. A run-away reflex pulls it back when Fear > Trust + Confidence. Fear comes from its ground sensors, Trust is adjusted in conversation with the player, and Confidence builds from experience.

FFWalkThePlank · UE5 · C++

Escape room: knowledge-gated puzzles

Object manipulation, code locks and weight puzzles. The robot can solve a puzzle only once it holds the prerequisite facts needed to understand it. Claude, ChatGPT and Qwen3-4B solve it unaided. Qwen3-1.7B needed one generic hint.

FFWalkThePlank · UE5 · Truth Engine

Maslow's Yard: learning by watching

No player input. The robot watches a smaller robot shelter from acid rain and follows it to cover. Then it watches one small robot free another that’s stuck, and later frees a stuck robot itself. This is observational altruism, running on a needs-driven architecture: LLM deliberation layered over learned reflexes, plan execution and motor control.

FFWalkThePlank · Maslow's Brain · UE5

I, Submarine: an LLM executive officer

The player captains an autonomous submarine whose executive officer is an LLM, acting through tool calls on the boat's systems. The game is about how trust builds between captain and crew. In progress.

AIXO · UE5 · LLM tool use

Delivery track record

Before AI, I spent decades designing and shipping software that had to work on the first try on fixed hardware: about 40 titles and products across consoles, PC, mobile, web and VR. A few examples follow; a full project list is available on request.

Multi-platform 3D rasterizer

An assembly-level shader compiler and software renderer that ran on many device families, with many complex sub-features. It was deployed in over 300 games.

Designed and built

Cross-platform C++ game framework

A framework that kept 98% of each game in shared C++ and 2% platform-specific, and shipped titles on iOS, Android and Windows.

Designed and built

Mobile MMO with 800,000 simulated enemies

Extended Ant Colony Optimization algorithms to drive 800,000 simulated enemies. The game design hid most of the effects of 3,000 ms server latency from players.

Designed and built

VR: multiplayer, instruments and combat

A multiplayer VR band simulator with custom guitar, keyboard and drum interfaces (HTC Vive), a moon-base defense game (Quest 3), and a space-combat client for an MMO (Oculus).

Designed and built

Branching-story engine with full undo

For a major publisher's iPad title, a plot flowchart that grew as the game was played, with full undo. It was an early version of the state graphs my AI work runs on today.

Built, in collaboration with the publisher's studio

Live-ops server for a console title

A Node.js game server and live-operations backend for a multiplayer console game.

Built, in collaboration with the game's studio

Background

Most AI products fail in the plumbing around the model, not in the model itself. I've spent decades building that kind of plumbing (engines, tools, networking, real-time systems), and I now apply it to LLMs.

I keep the whole system in view and work directly in the code: I design it, build it, measure it and write it up. No slides.

Working together

Remote from San Francisco. Hourly or fixed-scope. Engagements can start within days.

Engagements are contracted through Monkey Business, which has also built developer tools, music tools, networked classroom presentation systems, and games for mobile, VR, console and networked play. Project list on request.