Everything below is what has actually been built and shipped since December 2025, each system described by what it does and where it runs. One operator, a managed fleet of coding and research agents, and a self-hosted control plane that schedules, supervises, and audits their work.
Public and deployed work is shown in full. Private systems are described at an architecture level — no secrets, hostnames, or credentials.
Public operations shows a scrubbed health snapshot of the fleet behind the work, and says so plainly when the feed is down instead of showing stale numbers.
The three systems that carry the most weight: the agent infrastructure, the product proving ground, and a production data platform built end-to-end.
A self-hosted AI orchestration monorepo — a Go/Postgres scheduler, a Chrome MV3 browser-automation extension, a multi-agent fleet manager, and a React operator dashboard — that routes work across Claude, Codex, Gemini, and Antigravity CLI runners.
A full bid-to-closeout construction-operations platform for exterior cladding contractors, hardened from prototype to a deployed, pre-pilot DigitalOcean platform in roughly two months.
A production NHL intelligence platform built from scratch to live in one quarter (~13 weeks): advanced analytics, trade lineage, and a fantasy team-builder.
Agent infrastructure, products, learning tools, the attention substrate, a game engine, ontology research, and the fleet backbone.
The control plane, deliberately decomposed into standalone, versioned repos in a May 2026 extraction + governance pass.
Operator dashboard grown 0 → 48 tab views (~21k lines TSX): dispatch, fleet telemetry, kanban with live polling, research, quota ledger, ontology. Full TypeScript-strict rollout + boundary-doc enforcement.
The extracted multi-agent runtime: steerer fan-out, a pane-wakeup watchdog that detects and unsticks frozen tmux panes, a 10-hook lifecycle framework, OTLP telemetry, and account-credential multiplexing.
Standalone Fastify sidecar bridging a Chrome MV3 extension to mesh: tab-lease APIs, a capture-modal pipeline with HMAC ingest, per-dispatch research witnesses, and a 16-row version compatibility matrix.
Fleet deploy + topology orchestration: canonical topology for a 6-box fleet, an 11-repo release-train catalog, ordered 4-phase deploys, health monitoring with failover, and a portable package planner.
Meta-repo tying 11 private repos under one manifest + release gate + a 490-line suite-doctor emitting structured JSON health reports (EP-01..EP-19 governance plan).
Shared TypeScript contracts (capture envelope, observation events, HMAC headers) consumed by 9 repos — a single source of truth for cross-component data shapes.
Also: llm-accounts (47), capture (32), emacs-context (11), operator-sdk (9), research-pipeline (8).
Public and self-hosted product surfaces beyond the flagship platform.
This site: grew to 10 Canvas/WebGL ecosystem simulations (each on its own subdomain), a live multi-model debate surface, an ops dashboard, and a metamorphic visual-regression harness.
Self-hosted chess trainer built in 18 days: a pool of 4 warm Stockfish processes, multi-model LLM move coaching with an LRU cache, an adaptive Lichess-puzzle picker, and a sandboxed systemd service.
Self-hosted multi-user AI chat for a small trusted group (LibreChat + LiteLLM behind Cloudflare Access) with per-user spend caps and a subscription-ledger / route-decision API.
Early community-logistics co-op pilot (Next.js 16 + Firebase + Postgres) with a price-intelligence module to aggregate household demand.
Tools that turn reading, recall, and proof work into a tracked learning loop.
A native X11 HUD + Go service (~33,600 lines) that turns desktop activity into a real-time study coach: a target-run lifecycle engine, a proof-capture pipeline, and importable schema/client packages.
LLM-assisted math-proof visualization: parse → Lean 4 formal verification → SymPy computation check → Manim rendering, with a PDFium reader and Drive/Firestore sync. Concluded with a deliberate, clean public deprecation.
Standalone learning workbench: 11 routes (Today aggregator, cross-device PDF resume, spaced recall, glossary dependency graph, ontology atlas) with strict XState statechart discipline throughout.
A Personal Event Log substrate that correlates what the operator says, sees, and does into one queryable stream.
Always-on speech-to-text daemon writing HMAC-signed observations to mesh, an attention-window rendezvous protocol with a race-safe generation guard, and a voice-command reactor that fires local handlers.
Continuous multi-monitor capture with dual HLS hot/cold tracks, focus tracking, a host-local API, and a nightly AI archive-captioner running on a local GPU.
Physical-space tracking that fuses UWB positioning, IMU head orientation, and webcam face detection to infer which monitor has the operator's attention and drive window focus.
A strategy-game research artifact about projecting complex state faithfully.
An engine-first strategy game (TypeScript / PixiJS) where the engine owns all semantics and the UI is a pure projection: four rules profiles, fog-of-war commute invariants enforced in CI, an AI tournament balancer, and 10 watchable AI-vs-AI scenarios. Includes a 451-error strictness overhaul and a full a11y pass.
The vocabulary and belief layer that keeps the rest of the work coherent as it changes.
The operator knowledge substrate: a goal registry, a four-layer decision hierarchy, structured briefs for 8 sub-projects, and a research flywheel that evaluated 506+ topics in the window.
A 149-document corpus defining a layered ontology of digital life, with a typed four-rung abstraction backbone, a CLI slice-validator, and a formal olog fragment typing the event pipeline.
The infrastructure that makes a one-person, agent-assisted fleet operable.
GNU-stow dotfile management: profile-isolation-aware agent hooks, steerer autolaunch, capture/whisper systemd units, git push guards, and cross-agent session markers — the operational backbone of the whole fleet.
Custom Spacemacs layers for mesh integration: org-link types for mesh and proof surfaces, shell lifecycle event forwarding, attention-window binding, and a live key-coach HUD.
A clean, security-conscious first-party Chrome extension (playback speed, voice-repair EQ, per-site hotkeys, per-machine identity) with zero external network calls.
Each repository was analyzed with git log --since=2025-12-01: commit counts, date ranges, and the actual subsystems that shipped. Four cloned/forked repositories with no original engineering were checked and excluded. Some private repositories (financial records, scraping tools) are intentionally omitted entirely.
The same discipline — routing, auditability, review gates, and operator visibility — is what I help teams build into their own AI-assisted systems.