The skill set, in plain text

What I actually build.

Full-stack AI engineer: model layer, retrieval, agents, the interface on top, and the deployment that puts it on the internet. Every number below is measured off the running system, not estimated — where a figure couldn’t be verified against something on disk, it isn’t here.

LLM application engineering

  • RAG & hybrid retrieval
  • Agent orchestration
  • Evaluation harnesses
  • Tool-calling & streaming
  • Structured output, defensive parsing

Local & private inference

  • MLX on Apple Silicon
  • Quantized 35B-class models
  • OpenAI-compatible gateways
  • Model routing & A/B
  • GPU contention scheduling

Integration & data

  • Postgres + pgvector
  • NL→SQL with safety guards
  • PII detection & substitution
  • Multi-tenant isolation
  • Legacy platform bridges

Front end & real-time graphics

  • Next.js / React / TypeScript
  • Three.js & WebGL
  • Canvas 2D, PixiJS, Phaser
  • Scroll-scrubbed video engines
  • Responsive & mobile-hardened

Motion & generative video

  • Local diffusion video pipelines
  • Higgsfield / Draw Things CLI
  • Frame-locked seamless chaining
  • ffmpeg grade, encode, scrub-tune
  • GPU job queues & contention guards

Deployment & operations

  • Manifest-driven deploys
  • Cloudflare Pages & Tunnel
  • Docker, launchd, pm2
  • Managed services & watchdogs
  • Playwright QA in the pipeline

Selected work

The receipts.

Six systems, all built and running. The scroll above is these, as a place.

Long-term memory for an AI coding agent

Indexes every past session into Postgres, scores them with a local LLM judge, embeds the keepers, and injects the relevant ones back into new sessions automatically. Retrieval is vector search fused with full-text by Reciprocal Rank Fusion. A separate audit layer extracts testable claims from past answers, checks them against the filesystem, and flags the false ones on recall — the system catches its own hallucinations.

6,592 sessions 650,953 turns pgvector HNSW · 1024-dim RRF hybrid retrieval Python · Postgres · Ollama

Autonomous multi-agent coding loop

A local model works a backlog across parallel project lanes overnight while a frontier orchestrator reviews and promotes the work. Orchestrator and workers are strictly separated — workers verify but never write. A verdict only counts if it carries an exit code, a quoted log tail and a file reference, so unproven passes get demoted. Includes atomic task claiming with stale-claim recovery and three-strike escalation to a human.

2,638 commits 8 parallel lanes 9 weeks Evidence-gated verification

Multi-tenant AI layer for a tour company in New Zealand

A production platform where every AI task defaults to a local model, and only routes to a frontier provider when a tenant admin explicitly opts that task in. Before anything can leave the box, a local NER model running in an isolated subprocess finds and substitutes personal data, which is rehydrated on the way back. Natural-language reporting compiles to SQL behind a safety guard, with a stored-query fast path.

53 AI-backed API routes Consent-gated routing Local PII substitution Next.js · Postgres · ONNX

Private OpenAI-compatible inference service

A drop-in replacement for OpenAI in a production deployment: the legacy platform points at it unchanged. Four task-specific skills are routed by the requested model name, with no silent fallback on an unknown one. The SQL skill is grounded with few-shot examples retrieved by cosine similarity from an embedded corpus, and degrades gracefully if the vector store is missing. Zero third-party dependencies.

225 embedded query pairs 4 routed skills Zero dependencies Runs as a managed service

Model evaluation harnesses

Unattended bakeoffs that gate adoption on the things that actually break an agent loop. A candidate 35B model was disqualified outright for failing tool-calling before it ever reached production; survivors were throughput-benchmarked on identical hardware. Results include an honest limitations section and a documented negative result rather than a buried one.

Tool-call gating Same-hardware benchmarks Negative results published

Synthetic training-data pipeline

Generates natural-language→SQL training pairs by wrapping the live production system prompt and schema around a local 35B model, so every pair is in the exact dialect the real system expects. Generated SQL is validated against an in-memory database before it’s kept, and prompt changes are scored as pass-rate against a frozen validation set.

~9,000 validated pairs SQL-validated on generation Frozen-set scoring

Background

Thirty years of production systems.

Certified FileMaker developer and independent consultant. Before AI: student information systems for a K–12 board and a private school, a Fortune-500-facing developer role at one of the largest FileMaker premium partners in the world, e-commerce search during the dot-com boom, and a college instructor in dynamic web development. The reason the AI work lands in production is that the production part was never the new bit.

PythonTypeScriptNode Next.js / ReactPostgres + pgvector SQLiteMLXOllama ONNX RuntimeDockerCloudflare FileMakerPlaywrightffmpeg

The front end

Twenty-two live sites. Go and click them.

Every thumbnail below was captured from the live site today, and every one is a link. Cinematic scroll-worlds on a video engine I wrote, real-time 3D in Three.js, and games — all built, deployed and still up.

Cinematic scroll-worlds — pre-rendered camera flights, scrubbed by scroll

Thirty years, one continuous take Thirty years, one takeCareer fly-through Colosseum, 107 AD Colosseum, 107 ADHistorical world Pompeii PompeiiHistorical world Roman Triumph TriumphCanvas 2D + video siTOTis siTOTisCompany front door Worlds hub WorldsHub Papercraft diorama world Papercraft dioramaClient concept Isometric diorama world Isometric dioramaClient concept

Real-time 3D — Three.js / WebGL

The Quiet Wing The Quiet WingExplorable 3D gallery The Woods The WoodsProcedural 3D forest walk Museum of Play Museum of PlayWalkable games museum siTOTis Games arcade The ArcadeWalkable games portal

Games — Canvas 2D, one mechanic each

DRIFTDRIFTTwo buttons TETHERTETHERTwo buttons FATHOMFATHOMOne button VIGILVIGILChoose what to lose COUNTERPOINTCOUNTERPOINTSoundtrack is the score Dad GamesThe Card RoomNext.js party games

Product & editorial

AUTO_HERMES AUTO_HERMESAutonomous coding loop Five Hours Back Five Hours BackAI assessment offer Compliance audit Compliance auditData analysis & deck Fridge FridgeMicro-site

Motion

I build the video pipeline, not just the clip.

The scenes you scrolled through were generated, chained frame-to-frame so the camera never cuts, colour-graded per scene and encoded for scroll-scrubbing — then verified in a headless browser that measures every seam. Separately, I run a local diffusion-video rig on Apple Silicon with a job queue that arbitrates GPU time against a language-model workload on the same machine, including pause and resume at segment boundaries.

Frame-locked seamless chaining Per-scene ffmpeg grading Automated seam QA Single-GPU job scheduling Higgsfield · Draw Things · ffmpeg

Open to

Remote AI engineering.

Full-stack: I take AI work from the model layer through the interface and out to a live URL. Forward-deployed / solutions engineering, applied AI, or AI-heavy product work. Based in Ontario, Canada. Remote, and comfortable with it — everything above was built that way, alone, from this desk.

[email protected] The career story

The scenes above are AI-generated and colour-graded; the camera move is one continuous pre-rendered take, scrubbed by scroll. Built, and the numbers verified, by me.