beginner
ai-ops
demand-forecasting
rebalancing

Levy AI Ops Overview

Demand forecasting, rebalancing recommendations, and joint swap+rebalance routes — the AI layer on top of Levy Fleets.

Levy Fleets TeamMay 18, 20266 min read

Levy AI Ops Overview

Levy AI Ops turns Fleets from a "manage what already happened" platform into a "do what's about to be needed" platform. It is a three-capability product line that runs on the same Supabase ride data you already have:

  1. Demand forecasting — predicts how many rides each H3 hex will see in the next 1, 4, or 24 hours, conditioned on weather, events, and historical patterns.
  2. Rebalance recommender — compares the forecast to your current vehicle supply and proposes ROI-ranked moves: "Move N vehicles from hex A to hex B by T. Projected lift: $X.XX."
  3. Joint swap + rebalance routing — packages low-battery swap stops and rebalance moves into a single ordered route a technician follows in the operator-app.

Where you'll see it

SurfacePathWhat it does
Demand map/dashboard/analytics/heat-mapsH3-hex overlay with predicted demand, actual history, and unmet-demand heat
Rebalance/dashboard/operations/rebalanceRanked card grid of recommended moves with Accept / Snooze / Dismiss
Top-3 widget/dashboard/operationsDashboard widget showing today's three highest-lift moves
Operator-appRoute tabOrdered stop list for technicians, with offline-tolerant completion

How it's packaged

AI Ops is a paid add-on, not part of the base SaaS. Three tiers:

  • Starter — Phase 1 demand surface only.
  • Pro — Demand surface plus the rebalance recommender.
  • Enterprise — Everything plus joint swap+rebalance routes for technicians.

The active tier is controlled per subaccount by the ai_ops_tier flag and a few related settings. See Feature flags and tiers for the full list.

What's actually doing the prediction

Phase 1 ships with a LightGBM gradient-boosted regressor trained nightly on Modal, with a pure-TypeScript fallback model that runs in-process when Modal isn't configured. Predictions are stored in demand_forecasts and refreshed hourly. The recommender is closed-form math on top of forecasts — it isn't a separate ML model.

For routing, Phase 3 uses Google OR-Tools (a vehicle routing solver) by default, with Routific available as a higher-quality enterprise option. If neither is reachable, a local Clarke-Wright + 2-opt savings solver runs in Node and handles routes up to ~30 stops in under 100ms.

What it is not

  • It is not surge or dynamic pricing. AI Ops is an operations tool, not a pricing tool.
  • It does not auto-dispatch vehicles or auto-execute moves. Every recommendation is suggested — the operator (or technician) confirms.
  • It is not real-time. The smallest time bucket is one hour.

Next steps

Need help?

Questions about AI Ops? Email support@levyelectric.com.