CommandAGI (repo) — branchable execution environments for agents. I am building a live agent platform where computer and robot sessions are observable, interruptible, snapshotable, forkable, resumable, and metered. The system couples a Next.js/React PWA to a Hono control plane on Cloudflare Durable Objects, D1, and KV; keyless GCE provisioning; an in-VM Python computer-use runtime; data-driven 3D worlds; BYO-robot registration; Python, TypeScript, and Go SDKs; WebRTC intervention; finite-capacity scheduling; and prepaid settlement.

Canvas Engineering (docs, research article) — declared causal macrostructure compiled into diffusion-transformer computation. A typed schema specifies latent regions, directed connectivity, update clocks, and loss roles; the compiler lowers it into exact attention masks, loss weights, and frame mappings, making absent edges hard conditional independences. The evidence program, ROAM (dossier), spans 26 experiments and 236 runs: a frozen three-loop CogVideoX-2B graft reached 0.073 action loss with 350K trainable parameters versus 0.088 with 11.7M unfrozen, while three reasoning hypotheses failed. A 23-region/42-path cortical canvas matched a dense model at R² 0.825 vs. 0.826 using 19.6% of the connections and ~5× less training time; it predicts TRIBE encoder outputs, not measured biology.

WIN / B-WIN · SC-WBD (live results, dossier) — heterogeneous multirate world and brain dynamics. WIN retains each state variable's native support, frame, clock, units, uncertainty, and update law, connecting fully qualified (entity, domain, support, channel, clock) addresses through typed interactions and evidence-permission paths. SC-WBD instantiates the program over 414 regions and 26.3M parameters with declared forward operators and four public checkpoints. Checkpoint 003 improved density score by ~0.04 nats over VAR/AR baselines on 27 held-out participants; 004 did not reproduce it. Leave-one-source-out found 64-channel EEG at +0.0144 and fMRI at +0.0010, while simulator pretraining scored −0.0034 on measured data; all five biological claim gates currently read could_not_run, hence zero validated brain claims.

Limboid (repo, BoidKit, build archive) — centralized fluid power, distributed low-mass actuation. One water/glycol pressure source, accumulator, trunks, and zone manifolds drive parallel artificial-muscle bundles through inexpensive proportional valves; deterministic pressure/position/contact loops run at 100–200 Hz below Linux-class estimation and whole-body control. Full-body and upper-body rigs, valve manifolds, printed hand bones, tendon/soft-hand prototypes, CAD, and electronics exist; the centaur is under construction. Per-axis cost, force density, efficiency, and whole-robot BOM remain engineering targets, not measurements.

SynthUX (dataset, dossier) — grounded visual computer-use data without full VMs. A grammar expands goals into arbitrary-depth task trees whose terminal affordances execute as low-level mouse/keyboard events inside Windows, macOS, and browser simulators; screenshots, observed state, videos, generated artifacts, and node-level alignment remain causally connected. The v2 schema/world/resolver/scheduler/validator path has 106 tests and demonstrations with up to four actors, three environments, and cross-stream channel handoffs; v1 still generates the current public dataset.

Chem-0 (technical essay, follow-on) — inspectable low-cost robot control. An LLM operates a LeRobot SO-101/SO-100 through one typed tool surface shared by Electron and MCP. TypeScript owns sessions, SQLite events, artifacts, and model streaming; a JSON-lines bridge reaches Python/OpenCV/LeRobot, where calibrated joint/workspace bounds, step interpolation, URDF kinematics, and placo FK/IK gate commands before the Feetech bus.

EEG Acquisition Chain — cortical current dipoles to multiplexed ADC input. The deterministic interactive model spans volume conduction, CSF/skull/scalp spatial mixing, contact mechanics, CPE electrode impedance, AFE common-mode leakage, gain/filtering, and AMUX sampling. It separately exposes attenuation, irrecoverable spatial mixing, Johnson/amplifier/quantization noise, impedance imbalance, settling, charge injection, crosstalk, skew, and real alias folding, with every parameter marked by provenance and observability.

Earlier systems — mechanisms, not stack badges. BSBR explored chunk attention plus block retrieval for near-linear long-context inference; TensorCode encoded arbitrary Python objects through differentiable runtime code generation; MPNets unified heterogeneous learning rules on an executable graph; Node Tree generated paired language descriptions and environment demonstrations; MLN-Dashboard was a modular research-visualization server. Hardware work includes a verified ESP32-S3 UVC webcam, a reverse-engineered non-UVC microscope viewer, CNC machines, actuators, and embedded control.

Production systems. At AGI, Inc. I worked across the agent runtime, mobile integration, and eval tooling, then owned API/SDK/partner surfaces spanning iOS, on-device LLMs, quantization, schemas, and the agent control plane. Earlier I was founding engineer / software engineer #2 at Breezy AI, working across a Rails/Next.js AI-receptionist monorepo.