Six pillars.
One control plane.
HexaGrid unifies grid intelligence, price forecasting, quantum-hybrid scheduling, RL dispatch, carbon routing, and GPU telemetry into a single operational layer — driven by the QLLMe™ engine.
Each layer is independently functional. Together they form a closed-loop system: real-time data feeds the forecaster, the forecaster feeds the scheduler, the scheduler feeds the RL agent, and telemetry feeds back into every decision.
Live wholesale electricity prices from all five US ISO regions via the EIA API — cached with a 15-minute TTL and updated continuously. Grid volatility stops being noise and becomes a precise scheduling input.
CAISO · ERCOT · NYISO · ISONE · PJMA TensorFlow LSTM trained on two years of historical price data generates a 120-minute forward forecast in real time. That window is enough lead time to defer any flexible workload away from a price spike before it arrives.
LSTM · 1.92% MAPE · 120-min HorizonThe QLLMe™ engine runs three approaches in parallel: classical COBYLA/BFGS for low-latency deterministic decisions, QAOA variational quantum circuits for combinatorial scheduling, and PPO reinforcement learning for adaptive dispatch. The best decision wins.
QLLMe™ · QAOA · Cirq · QPU-ReadyA Stable Baselines 3 PPO agent trained over 200,000 steps learns from live market behavior — not simulated price curves. Benchmarked at 15–30% cost reduction versus flat-rate scheduling in live grid conditions. Savings compound as the agent adapts.
PPO · SB3 · 200k Steps · 15–30% ReductionLive carbon intensity from Electricity Maps across all five ISO regions. The same GPU workload can produce 3× more CO₂ depending entirely on when and where it runs. HexaGrid makes that difference visible and automatically routes against it.
Electricity Maps · Scope 2 · Up to 50% CO₂NVML telemetry polls every GPU every 10 seconds — temperature, power draw, VRAM utilisation, ECC error counts, fan speed. An IsolationForest anomaly detector catches combined-metric failure signatures before any single threshold would fire an alert.
NVML · IsolationForest · 0–100 Health ScoreRL dispatch vs flat-rate
carbon-first routing
LSTM model
simultaneous feeds
Three optimization layers.
One unified decision.
QLLMe™ is what separates HexaGrid from conventional energy management tools. Rather than committing to a single algorithm, it runs three complementary approaches in parallel and selects the optimal decision per workload in real time.
Cost, carbon, capacity, and PUE are resolved simultaneously as a multi-objective problem — not sequentially as competing constraints. No manual trade-off tuning required.
As quantum hardware matures, the QAOA layer migrates to real QPU backends via IBM Quantum or IonQ — with zero code changes to the rest of the platform.
Deploys in 15 minutes.
HexaGrid runs on any NVIDIA GPU infrastructure — no new hardware required, no vendor lock-in. A single uvicorn process, a static dashboard, two API keys.