Foundation models, fused into one understanding.

01

Time-series forecasting

TimesFM 2.5 is the workhorse for zero-shot load, solar, wind, EV charging, and heating demand forecasts. Where context windows or quantiles matter, we ensemble with Chronos and Moirai. For interpretable, multi-horizon feeder load forecasts the platform falls back to Temporal Fusion Transformers — the operator gets a per-feature attribution alongside the point forecast. On consumer hardware, top-performing time-series foundation models in 2026 have been measured at ~47% MASE reduction over seasonal naïve.

02

Vision

Qwen3-VL handles substation, line, and ROW inspection at the cloud edge. Pixtral runs the same workflow on a Hailo-10H or Jetson Orin Nano Super at line-power budgets, suitable for pole-mount and substation-cabinet deployment. Both produce structured findings — defect class, severity, location — that flow into the work-order system without a human in the loop unless severity exceeds a threshold.

03

Acoustic anomaly detection

PANNs CNN14 embeddings + a per-asset linear classifier. Fingerprints are enrolled once at commissioning. The platform watches each transformer's signature continuously and flags deviations: partial discharge, corona, mechanical fault, fan failure. Edge inference on YAMNet means raw audio never leaves the perimeter.

04

Embeddings & memory

Gemini Embedding 2 in Matryoshka 1536-d for the workspace tier. Every site under management has a long-term memory — events, anomalies, work orders, post-mortems, agent decisions — that is semantically searchable and that the agent retrieves from before it acts.

05

Reasoning

Gemini 3.1 Flash Lite as the conversational interface and tool-orchestrating agent. Tool calls are typed (zod-validated) and per-tenant; the model never invents a tool it wasn't given. Where heavier reasoning is required (long planning horizons, ambiguous policy decisions), the platform escalates to a larger reasoning model with the same tool surface.

06

Provenance, end-to-end

Every prediction the platform makes is stored with the model name, version, weights hash, input window, and confidence interval. Operators can replay a forecast a year later with the exact same inputs and verify the result is identical. This is the same audit-trail discipline coast applies to environmental predictions; the grid version adds the additional layer of regulator-facing decision logs for autonomous dispatch.