Forecast Substation A-12 load tomorrow at 15-minute granularity. Highlight peak risk.
Tomorrow's peak forecast is 186.4 MW at 17:30 UTC. Ensemble disagreement ±3.4%.
Top driver: lag_24h_mw (SHAP +16.2 MW). Exceedance prob > 195 MW: 18.4%.
LLMs draft a structured UserTaskPlan. You approve, edit, or reject before any tool runs. No silent decisions.
Every quantitative claim is verified against the session's tool outputs. Unsourced numbers are stripped or flagged.
Statistical baselines, ML, deep learning, and time-series foundation models — picked by validated CV metrics, not vibes.
Recommendations against EMS/BMS/DER systems route through governed approval. No autonomous control in any release.
Tenant-scoped resources for Postgres rows, ClickHouse partitions, Redis keys, MLflow runs, and BentoML caches.
Full responsive web copilot plus an installable PWA for live dashboards, alerts, and one-tap approvals on call.
Two surfaces, one workbench
The web copilot is for deep analysis: NL chat, plan review, forecast studio, explainability, governance. The PWA is for on-call ops: at-a-glance forecasts, swipe-to-acknowledge alerts, one-tap approvals.
- • Tenant onboarding wizard (5 steps)
- • NL query → plan-review → grounded report
- • Forecast Studio with leaderboard + conformal intervals
- • 7-detector anomaly investigation with context window
- • SHAP waterfall + counterfactual forward & inverse
- • Governance: approval queue + audit + Merkle proof
- • PWA: live dashboard, swipe alerts, one-tap approvals