Our lab, where we try things before clients ask
This isn't client work. It's where we run the latest open models on our own hardware, operate Ethereum validators and build research prototypes, so our advice comes from hands-on experience.
- Hardware
- 4× RTX 3090 · 96 GB VRAM
- Models
- Qwen · DeepSeek · Liquid AI
- Research
- Vision · peer-to-peer data
- Ethereum
- Validating since Dec 2020
AI workstation
Open-weight models on our own GPUs.
We run open-weight models locally to learn what each new release can really do, for text, code, images and video. Nothing here goes to an outside AI provider. When a client asks for private AI, we recommend setups we've already run ourselves.
The machine
- GPUs
- 4× NVIDIA RTX 3090, 96 GB VRAM
- CPU
- AMD Threadripper 3960X, 24 cores
- Memory
- 256 GB DDR4
- Models
- Qwen, DeepSeek, Liquid AI and others
- Workloads
- Text, code, images, video
We set up the same kind of system for clients, in their office or an EU data centre. See private AI →
- Qwen
- DeepSeek
- Liquid AI
- llama.cpp
- CUDA
Ethereum node & validators
Validating since the Beacon Chain's genesis on 1 December 2020.
We've run our own Ethereum validators since day one of the Beacon Chain, alongside a go-ethereum archive node on our own hardware. It started with our contribution to the ETH2 Staking Launchpad in 2020 and keeps us hands-on with client releases, hard forks and the day-to-day work of keeping a validator online.
What we run
- A go-ethereum (geth) archive node on our own hardware: Ethereum's full state history, queryable locally.
- Prysm beacon node and validator client on the Beacon Chain, running since genesis.
- Kept up to date through every network upgrade since 2020, including the Merge in 2022.
- go-ethereum
- Prysm
- Archive node
- Beacon Chain
- Validators
Cortical Mirror
A vision model inspired by the brain's visual cortex, built for a hackathon of the California Institute for Machine Consciousness.
We built Cortical Mirror for a hackathon run by the California Institute for Machine Consciousness (CIMC). It learns online from a live webcam using only local learning rules, the way predictive-coding theories describe the visual cortex, to see how far brain-like learning gets without a pre-trained network.
What we built
- Learns motion-sensitive cells from the live video stream.
- Finds objects and builds a scene graph of what it sees.
- Tested on the UCF101 action-recognition video dataset.
- Python
- GPU
- Predictive coding
- UCF101
ph-reactor
An experimental peer-to-peer document store in Rust.
A native Rust take on Powerhouse's document model: every change is an event, signed by its author and synced between peers without a central server.
What we built
- Event-sourced storage with a write-ahead log and snapshots.
- Per-field last-writer-wins merging with vector clocks.
- Operations signed with ed25519.
- Peer-to-peer sync over libp2p, with gossipsub plus a custom request/response protocol.
- Rust
- libp2p
- ed25519
- Event sourcing
Workshop
A 3D printer with laser module for hardware prototypes.
Not everything we build is software. With our Bambu Lab H2C 3D printer and its laser module we prototype hardware in-house: enclosures, mounts and parts for our own projects, printed or laser-cut the same day.
What we use it for
- Enclosures and mounts for servers, nodes and devices.
- Quick hardware prototypes to test an idea before ordering parts.
- Laser engraving and cutting for small parts and labels.
- Bambu Lab H2C
- 3D printing
- Laser module
- Prototyping
Want private AI on your own servers?
See private AI →