Skip to content
Brad Stancel

The lab

Most of what I know, I learned by building something real. This is where that happens: the AI hardware in my office and the systems running my house.

Running today Being built, tested, or planned

AI lab

I run language models on my own hardware for the same reason I run my own servers: to understand how the whole system behaves, not just the model. What I learn here goes straight into the private AI and automation work I do for businesses.

Related writing

  • Two-node NVIDIA DGX Spark cluster

    Running

    Two Grace-Blackwell GB10 machines joined by a 200 Gb RoCE link, serving large models with tensor parallelism across both nodes.

    • Stock NCCL deadlocked on the new chips. Building it from source for the right architecture fixed it and more than doubled throughput (2.4×).
    • Fine-tuning and quantization experiments for smaller, task-specific models.

  • Coding-agent research

    Ongoing

    Why does the same model do better work in one coding harness than another? I test agents on real repositories and on an arcade-game build where a black screen can't be argued with.

    • A two-model workflow: a frontier cloud model writes the specification, a local coding model implements it.
    • Focus on context, tool feedback, sub-agents, and how an agent verifies its own work.

  • One gateway for AI tools

    In daily use

    More than 25 Model Context Protocol servers behind a self-hosted API gateway, with a custom plugin that exposes local tools to any model I'm working with.

  • Multi-agent reporting pipeline

    Active build

    An orchestrator hands a plain-English report request to specialist agents for requirements, knowledge retrieval, SQL drafting, and validation, passing work through a shared artifact store.

  • Code knowledge graph

    In use

    A graph of a very large production codebase (Neo4j with GraphRAG tooling) that agents and people query before changing anything.

  • Local voice assistant

    Running

    Speech recognition, a local language model, and speech synthesis for the house, with no cloud service in the loop.

HVAC and home systems

I'm self-taught on HVAC, and I got good at it the way I learn everything: by measuring my own systems until the behavior made sense. The house has two HVAC systems with electric heat strips and electric water heating, so small faults show up as large bills.

I'm not a licensed HVAC, electrical, or solar contractor. Everything here is my own home. For refrigerant work or anything your local code regulates, hire a licensed professional.

Related writing

  • HVAC diagnostics

    Ongoing

    Tracking down why the systems use what they use. A diagnostic pass on the electrical system during the solar build found heat strips running when they shouldn't have been, a fault that was quietly costing hundreds of dollars a month.

  • Home automation

    Running

    Home Assistant with Z-Wave and Zigbee devices, energy monitoring, and local control, so the house keeps working when the internet doesn't.

  • Storage and backups

    Running

    A ZFS-based NAS with snapshots, replication, and backups that are checked by restoring files, not by trusting a green status.

Solar, batteries, and off-grid

Energy systems are infrastructure with physics attached: the connector fits or it doesn't, and the voltage is right or it isn't. I designed the house system around a time-of-use rate plan, then kept measuring until the dashboards and the meter agreed.

Related writing

  • Solar and battery backup

    Peak shaving daily

    An EG4 FlexBoss 21 inverter and GridBoss distribution, six 51.2 V LiFePO4 batteries (30.7 kWh), and 450 W Canadian Solar panels.

    • Charges during super-off-peak hours and covers the expensive morning windows.
    • Lessons the manual left out: an undocumented sync step, and batteries reporting 99% charge while actually sitting at 50 to 70% until they were cycled and balanced.
    • Next: the rest of the panels, generator integration, and more storage.

  • Battery telemetry

    In progress

    Reading battery-management data over CAN and RS-485 with small Linux computers, so battery health becomes logs and alerts instead of a number on a screen.

  • Off-grid mesh network

    Running

    A three-node MeshCore radio network on RAK4631 hardware, with a solar-powered rooftop repeater, for messaging when the internet and cell towers are down.

Making things

If a machine is worth keeping and the part no longer exists, the answer should be to make the part. I'm building toward exactly that, with an open-source toolchain and no cloud-dependent firmware anywhere in the chain.

  • Parts manufacturing

    Planning, staged

    A capability for out-of-production automotive and tractor parts: a Tormach 1100MX for steel and titanium, high-temperature 3D printing for engineering polymers, and in-house circuit boards.

    • Toolchain: FreeCAD, OrcaSlicer, Klipper, and LinuxCNC, with KiCad for electronics.
    • Bought in stages so each capability comes online on its own.

  • Software-defined radio

    Experimenting

    Receive-only SDR experiments with antennas and Linux-based collection, to understand how modern wireless systems behave.

  • Video capture pipeline

    Running

    A headless capture appliance that records, organizes, and uploads video automatically, moved from a Raspberry Pi to a small x86 machine when it outgrew the Pi.

Want this kind of curiosity pointed at your business?

The same habits that fix a heat strip or a GPU deadlock work on reporting, automation, and infrastructure problems.

Start a conversation