The lab

Where hardware meets software. I did electronics and robotics as a teenager, then stopped when I left France in 2008. I picked the hobby back up in 2024 but never found the time. Since moving to a new place in Hong Kong with a much better lab, I'm trying again: even with two kids, AI now does enough of the heavy lifting that I can finally build hardware again.

Bigger models run here too, so the data doesn't have to leave home.

Inference cluster

Four NVIDIA DGX Spark units (GB10, 128 GB unified memory each). They sit on 10GbE into the house network and talk to each other over 200G QSFP while a proper cluster switch is on the way. The serving stack is still settling as they come online — this page will name it once it's boring and reliable.

What runs on it

Mainly my own coding models for Farmslot, and a place to compare open-weight models running fully offline.

The Sparks also take the heavier local-inference jobs behind EchoBridge: transcription, diarization and speaker embeddings, so meeting audio stays on hardware I control. Lighter EchoBridge services already run on a home GPU workstation. On-device work (Soothly, phone STT via Audiolab) stays on the phone; the lab only takes what phones can't.

Supporting machines

• Two home microk8s clusters that host different workloads (API workers, auth, ingress, monitoring, and similar).
• A desktop with an RTX 4080 for always-on GPU services (EchoBridge workers today; stays useful once the Sparks are up).
• Two 128 GB Apple Silicon laptops for parallel agent jobs (Farmslot slots and heavy local builds).
• A Mac mini as a small always-on edge box.

Hardware bench

• Embedded audio path for Soothly: ESP32 and Raspberry Pi 5, Pi audio HAT, cameras, and the bits needed to move cry classification off the phone onto a monitor.
• Capture: USB and lav mics used when testing Audiolab and EchoBridge.
• Fabrication: Bambu Lab A1 Combo, mostly for electronics enclosures and brackets, plus a soldering / hot-air rework station.
• Radio side-quest: bladeRF 2.0 micro for SDR experiments (notes to come).

Lessons / notes

Notes on bring-up, power and local STT will land on /notes once the cluster is boring.