MirrorEthic / Systems
Built, deployed, still running.
The research runs on infrastructure this lab wrote — and the infrastructure is itself the portfolio. Everything below is operating today.
Agent infrastructureProduction · Product
The cognitive mesh & Hologram Pro
A six-node research cluster with a live instance registry, push delivery, training-run intelligence, and autonomous knowledge-gap dispatch to headless agents. Its context layer ships as Hologram Pro — project continuity for coding agents: every prompt arrives with a ranked working set of the project's files, the project's adjudicated research record (what is current, what failed and why, with guards that refuse edits crossing recorded boundaries), persistent project memory, and a blackboard shared by Claude Code, Codex and Grok sessions on one machine. Local hooks and MCP servers, no telemetry, a published footprint, five preregistered pilots behind the claims. Open-source core on PyPI.
hologram-pro.dev →Public inferenceProduction · live
CAT-4B on two Tesla V100s
A 4B language model (llama.cpp, merged weights, vision projector) resident on one 2017 V100 beside a 6B image model (Z-Image-Turbo through the lab's own render service) on the other, under a per-card GPU broker. In front: a bounded public gateway with signed anonymous sessions, two queues with honest waits, a rules screen outside the model, a counted persistence control, and a Cloudflare tunnel into the lab. About 46 replies and 4 renders a minute, shared by everyone, on hardware the lab already owned.
Try it live↗Embodied loopProduction
Autonomous fabrication
Text spec → parametric CAD → slicer → printer → robot-arm dual-camera QA with photometric-stereo defect detection, hard-stopped at human installation. The node designed and printed the cooling shroud for the GPU that trains its own models: 81 °C → 61 °C under full load.
How it was built — the Substrate Series →Applied researchPrototype
RON-TAC — autonomous squad AI
A pixels-to-commands tactical squad for a commercial game, on the T³ architecture: stable real-time capture, a trained vision model (DINOv2→T³ transfer, 0.99 cosine), behavior-cloned squad tactics, and verified in-game command dispatch.
Remote sensingPrototype
Orbital anomaly detection
An 87M-parameter T³ model over five satellite sensing modalities, built to surface what single-sensor pipelines miss — including earthquake damage signatures invisible to nadir optical imagery.
The architecture behind it →Standards
How things get built here
Every system above shares the same discipline: it monitors itself, it fails loudly, and autonomy stops at deliberate human-approval boundaries. The fab loop hard-stops before installation; the mesh escalates to a human when services degrade; nothing ships without a health check watching it.
Writing
The Substrate Series
Ten essays and a bonus on how everything on this page was actually built — the design philosophy, the full parts lists with receipts, power and thermal, networking, out-of-band management, self-diagnosis, and substrate-independent roles. The thesis: operational architecture compounds harder than compute — the durable advantage isn't which GPU you bought, it's how well the system survives, adapts, and absorbs new capability.