Applied AI engineer with deep finance roots — I design and ship agentic, judge-gated, cost-aware systems, grounded in real financial work.
Bay Area · open to remote and relocation
Autonomy
Alpha4Gate plays ranked StarCraft II as a rule-based + PPO-policy Protoss bot with a Claude advisor riding shotgun, then closes the loop on itself: it drafts, tests, and validates its own code fixes to raise its own win rate. A themed two-pane self-play viewer streams two bot generations battling live; a 6-tab React dashboard shows the same game over the network.
Factory
void_furnace is a dark-factory autonomous coding pipeline: a planner, a coder, and a held-out critic run on a systemd timer, turning operator-filed GitHub issues into merged pull requests on a separate target repo with no human in the loop. The critic's prompt is built from the issue body, the diff, and the target's own governance docs -- structurally excluded from coder narration, so persuasive PR writing cannot buy a pass. void_furnace_calc, an arbitrary-precision Decimal calculator, is the target product it builds.
Orchestration
skill-mesh is a provider-neutral collection of roughly 50 agent skills spanning planning, building, reviewing, and shipping software -- each with one shared behavior contract and per-host adapters, so the same skill runs unmodified on Claude Code or GitHub Copilot. An 8-rail routing web sorts every operator fragment (a bug report, a half-formed idea, a status question) onto a pipeline rail without inventing new work; the core pipeline chains plan-init through build-phase to repo-update behind quality gates.
| Category | Name | Description |
|---|---|---|
| Improvement | b2_project_goblin | A local-first AI second brain that reads a project's real state, ranks improvement suggestions, and renders one into a /goal or plan seed. |
| Verification | changed-check | Maps a Git change set to the narrowest repository-declared validation commands, showing the full plan and rationale before running them. |
| Rigor | citation-needed | Gives each design choice in LLM-facing files a research-literature citation trail and an anchored 0-100 score, persisting vetted citations. |
| Management | dev-observatory | The workspace control plane: aggregates every project's goal/built/planned status into a CLI table, static HTML, or a local dashboard. |
| Retrieval | find-again | Builds a local SQLite/FTS5 full-text index over configured dev-memory artifacts and returns ranked matches with openable path:line locators. |
| Coordination | heads-up | Advisory, expiring claims over repo paths/state so parallel sessions see overlapping intent before colliding, via an append-only ledger. |
| Benchmarking | measure-twice | A local benchmarking toolkit and mt CLI that sweeps suites across a model roster, scores them, and keeps an evidence ledger for tier claims. |
| Measurement | mesh-lens | Normalizes Skill Mesh dispatch telemetry and outcome evidence into comparable cohorts by skill/model/project/task, without touching routing. |
| Aesthetics | on-brand | Compiles a project's DTCG design-value file into CSS/Tailwind/TS/Python/SVG brand outputs, and drafts a brand guide from any website URL. |
| Decisions | paper-trail | Durable, immutable decision records with explicit supersession -- one decision per YAML-frontmatter Markdown file plus a derived JSON index. |
| Consistency | same-page | Extracts narrowly defined status claims from authoritative dev artifacts and reports deterministic contradictions with exact source spans. |
| Routing | switchboard | A stdlib-only library/CLI that routes cheap judge/grader calls to a local model and defers to Claude on any failure, inert until configured. |
| Safety | tripwire | A read-only preflight CLI that turns known workspace failure modes into fast, evidence-bearing checks over repo state and command text. |
Fabrication
Turns a photo or text description into three LEGO build suggestions with 3D previews, parts lists, and a step-by-step instruction PDF.
Playful
A tiny cwp CLI that runs a YouTube channel's episode lifecycle plus the Pantsless Build loop: kid's voice to a kid-usable AI-built toy.
Storytelling
A markdown-first creative-writing workshop: shared AI prompt skills and a stdlib-only ss CLI that scaffolds standalone story projects.
Play
A local-first home AI that listens for play opportunities and runs parent-approved activities on a child iPad kiosk -- all on-device.
Wellbeing
Turns live NWS weather, air quality, and estimated pavement temperature into one green/yellow/red 'can I walk my dog now?' verdict.
Discovery
A for-fun xmarks CLI cataloging famous lost treasures with citation-gated facts and Bayesian-search scoring, exported as a static dossier.
Forecasting
A local, always-on portfolio engine: ingest, forecast, Black-Litterman blend, constrained optimization, with a per-stage dashboard.
Learning
Each topic folder is a hands-on learning ramp of runnable notebooks: 70% hands-on, 20% mentoring, 10% coursework.
Finance
A reusable Python CLI finance toolkit for a PTA/booster club that treats a Google Sheet as its database and generates monthly HTML reports.
| Project | Description |
|---|---|
| agora | A decentralized AI compute collective where contributors donate GPU compute to train and run a shared LLM. |
| tinstar | A canvas workspace for managing many concurrent Claude Code sessions as live-terminal widgets on an infinite spatial canvas. Added a per-session model-routing switchboard override, a model-attribution widget, tmux-selection-to-clipboard copy, and the Windows/WSL2 setup guide (9 aberobison commits). |
The projects above aren't separate experiments — they're one system, built and kept honest with a shared discipline: anything that claims to work has to prove it through a gate it can't pass by luck, and no tool grades its own work. It reads in three layers — the flagships that prove the hard claims, the utilities that keep them honest, and the wider range built the same way.
I'm an applied AI engineer with deep finance roots. I currently build AI and data tooling for private-asset valuation as an Applied AI Scientist (contract) at FEV. Before that I spent 10+ years in quantitative finance — Senior Quantitative Analyst, Treasury at Google; Senior Quantitative Research Analyst at Russell Investments; and earlier work at BlackRock — shipping applied-ML and forecasting systems into large-scale finance operations.
The way I work is a pipeline with quality gates: plan one step at a time, build each in isolation, review with independent adversarial passes, and only then ship. Most of what's on this site was built that way — autonomous, judge-gated, cost-aware — and grounded in real financial work.
I hold an MA in Economics and a Graduate Certificate in Computational Finance from the University of Washington, plus a BS in Computer Science and a BA in Mathematics. I'm Google Cloud Certified — Generative AI Leader.
Essays on the engineering behind these systems — self-improving loops, honest measurement, and the failures that taught me the rules.
The first time my StarCraft bot beat the game's AI, the pipeline built to detect wins flagged it — and eleven other games — as crashed. A story about producer-consumer drift: the bug class that lives in the relationship between two correct components, invisible to 682 passing tests.
In draft
A two-page summary of my background, roles, and the systems I have shipped.
Download résumé (PDF)Last updated
The best way to reach me is email — I read and reply to everything.
aberobison@gmail.com