GridMind

An autonomous infrastructure router that treats cloud compute like a high-frequency energy arbitrage market, dynamically shifting workloads to the cheapest, cleanest data centers in O(1) time.

Comment

Inspiration

Energy costs represent 30-40% of AI infrastructure spend at the hyperscale companies. Yet workload routing is static; we route code and models to regions without considering the actual physical state of the grid. Electricity wholesale prices swing 10x intraday, and cooling efficiency is thermodynamically linked to ambient temperature; even the materials used to produce power are not the same across the country. California runs 100% renewels during solar peaks; Texas burns coal at night, meanwhile, nobody is optimising for this. We asked, What if routing decisions were made in real time, physics-aware, and carbon-conscious?

What it Does

Ingests live grid data from 3 US datacenter regions (San Jose/CAISO, Ashburn/PJM, Austin/ERCOT): wholesale prices, carbon intensity, 24-hour forecasts, and ambient temperature. Scores regions on normalized cost, facility efficiency (PUE), carbon, and latency. Presets for Training (carbon-first), Inference (latency-first), Batch (cost-first). Claude's agent decides which region to route to—now or defer to a cleaner hour—with reasoning backed by actual numbers. The model judges; C++ does the math. Takes real action: Opens a GitHub PR with a Kubernetes manifest pinned to the chosen region, or boots a real Fly.io machine in that region. Verifiable. Auto-cleaned. Fleet autopilot routes entire job queues autonomously, accumulating savings with Claude-written explanations.

How we built it

Layer 1: Data Ingestion (Node.js/Next.js)

  • Parallel API calls via Promise.allSettled()
  • 3-second timeouts per API
  • Graceful fallbacks to historical data

Layer 2: Optimization Engine (C++ → WebAssembly)

  • Priority queue for O(1) optimal region lookup
  • Thermodynamic PUE scaling based on live temperature
  • Compiled via Emscripten: 10x faster than JavaScript

Layer 3: Autonomous Agent (Claude Sonnet 4.6)

  • Tool-use pattern with 5 callable functions
  • Claude autonomously chains tools and makes routing decisions
  • Explains reasoning and acknowledges risks

Layer 4: Dashboard (React/Tailwind/Recharts)

  • Real-time region matrix, cost calculator, carbon forecast
  • Claude's full analysis with risk warnings

*Layer 5: MCP * -fully built MCP allowing calls from agents and AI, fully autonomous workflow

Challenges we ran into

  • Finding real price data: Our first source (EIA) returns demand in MWH, not price, so we switched to GridStatus.io real-time LMP/SPP feed.

  • Score normalizations: carbon intensity (hundreds) initially swamped price and PUE, so weights and presents didn't change the ranking until we normalized every factor onto a common scale.

  • Making the agent fast: a multi-step tool chain, extended thinking took 30-40 seconds. We preloaded all the data into one forced tool call and dropped to 7 seconds while not losing any decision quality.

  • Free tier rate limits: GridStatus 1-req/s + monthly quota forced sequential cached fetching.

Accomplishments that we're proud of

  • Built a full loop system: not just a dashboard that recommends, but an agent that decides within guardrails and provisions real compute in the region that the math determined was the most optimal.

  • A real Fly.io machine booting in the agent-selected region.

What we learned

  • LLMs can't do it all, using an LLM or agent where it would be truly beneficial and not just to say we have it in our tech stack. -Guardrails belong in code, not in a prompt: filter the choice before the model -Context is king: pre-loading data into one call is dramatically faster than letting the agent fetch it step by step

    What's next f3or Gridmind

  • Bring your own cloud: each enterprise connects their own system so the agent can route into the customer's infrastructure ( Slurm, AWS/GCP), not a pooled account)

Built With

Share this project:

Updates