Inspiration

I wanted to answer a simple question: If a CFO has 30 seconds, can they tell where to reduce GPU spending—and what could go wrong? Low GPU usage alone doesn’t justify a budget cut. I built Reclaim to connect technical evidence to a practical financial decision: what to investigate, who to involve, and how to test a change safely.

What it does

Reclaim turns telemetry from a real 225-machine cluster into a dashboard for GPU spending decisions. It helps users:

  • Identify workloads worth testing on CPUs and low-activity sessions worth reviewing.
  • Estimate the value of freed GPU time and the potential cost of disrupting useful work.
  • Trace recommendations back to individual jobs and GPU measurements.
  • Find the researchers to consult first.
  • Download a trial brief with an owner, review date, spending limit, and stop conditions. It also separates freed capacity from actual bill reductions. Using fewer GPU-hours only saves cash when the organization can reduce a corresponding expense.

How I built it

I built Reclaim as a solo project using React and TypeScript for the dashboard, Python and FastAPI for the analysis and API, and Docker for reproducible setup. I used the provided telemetry, investigated the proposed findings, and implemented calculations that avoid counting the same opportunity twice. Optional Featherless explanations use evidence retrieved through MCP tools, while financial calculations remain in code. I used AI coding assistance throughout development and checked the results with automated tests, source-record comparisons, and browser testing.

Challenges I ran into

The hardest part was deciding what the data actually supported. A failed job doesn’t necessarily mean broken hardware, and low utilization doesn’t prove that a GPU can be removed. I also had to translate detailed analysis into language a CFO could understand quickly, while keeping assumptions and supporting evidence accessible.

Accomplishments that I’m proud of

I built a working path from a recommendation to its evidence and then to a concrete trial plan. One useful finding was that three researcher accounts represented 52.4% of the historical GPU-hours eligible for CPU-placement testing. That gives the team a focused starting point for conversations and experiments. The project also passed 29 automated tests covering accounting, investigations, explanations, and trial planning.

What I learned

The most useful output isn’t the largest savings estimate. It’s a recommendation someone can investigate, understand, and act on. I learned to distinguish observations from assumptions, capacity value from cash savings, and a promising trial from a proven result.

What’s next for Reclaim

Next, I want to connect billing and contract data, run small trials with workload owners, and compare predicted benefits with measured results. That would let Reclaim move from identifying opportunities to verifying which changes actually reduce spending without harming useful work.

Built With

Share this project:

Updates