LANtern

LANtern turns your office's idle GPUs into a private compute pool. Submit a job from any machine; it runs on real CUDA on hardware you already own before renting cloud. GPT-5.6 vets every job first

Comment

Inspiration

I was thinking of different ideas for this hackathon. I then realise that each time I am trying to train a ML model, I would use Kaggle's free compute, leaving my device's GPU unused. I thought of how I could avoid this wastage and thought of the idea to pool them together.

What it does

Pools idling devices' GPUs together into one shared pool and allows users who needs them to use it instead of having to rely on cloud compute which can be very pricey. This effectively reduces wastage of compute that are already available to most companies.

How we built it

I built it using codex, GPT5.6, Fable 5 and Opus 4.8 claude code.

Challenges we ran into

  • Real GPU sharing forces shipping the entire CUDA/PyTorch runtime — an unavoidable 2.5GB per-contributor payload.
  • Windows PyPI torch is CPU-only; our agent silently couldn't touch the GPU until repinned to CUDA.
  • Running users' arbitrary Python on donor machines unsandboxed risks stolen secrets, SSH keys, data exfiltration.

Accomplishments that we're proud of

Successfully allowing multiple contributors to contribute to the GPU pool and have a user make use of the compute in the GPU pool.

What we learned

  • Shipping real GPU compute is less about code, more about wrestling multi-GB CUDA runtimes across OSes.
  • CI-green never means it runs — a glibc mismatch and CPU-only torch both slipped straight past CI.
  • An LLM gatekeeper turns an unsafe-by-design feature into a defensible, honest trust layer — GPT-5.6 earns its place.

What's next for LANtern

Ensure security when transferring data. Allowing only authorised teams/users to use the pooled GPU. Allow GPU pool to work with macOS GPU, AMD GPUs and NVIDIA GPUs, and concurrently in the same pool.

Built With

Share this project:

Updates