Matching the right EV to the right rider on NCR's chaotic road network is a hard optimisation problem. A look under the hood of our dispatch engine.
Dispatch looks simple from the back seat: a car shows up. Behind that is a continuous optimisation running across the entire fleet — balancing battery state of charge, live traffic, driver shift windows and predicted demand. When we started measuring pickup time end-to-end — from booking confirmed to chauffeur at the kerb — our median was 11 minutes. Today it's 6.8. That 38% improvement didn't come from asking drivers to drive faster. It came from smarter matching.
Why NCR breaks naive dispatch
Most ride-matching algorithms assume a city behaves like a grid: nearest available driver wins. NCR doesn't work that way. A vehicle five kilometres away on Golf Course Road may arrive before one two kilometres away if the closer car is stuck at the Sikanderpur bottleneck. One-way flyovers, U-turn restrictions, and time-of-day lane reversals mean straight-line distance is a poor proxy for actual arrival time.
Route AI was built around this reality. Instead of optimising for distance, we optimise for predicted arrival time — using live traffic feeds, historical corridor speeds by hour and day, and a growing map of NCR-specific quirks that no generic routing API knows about. The MG Block U-turn at 6pm. The DLF Phase 2 service-road shortcut that only works eastbound. These aren't edge cases here; they're the commute.
State of charge is a first-class signal
Unlike fuel, charge can't be topped up in two minutes. Our matcher treats remaining range as a hard constraint, never assigning a vehicle that can't comfortably complete a trip plus its return to the nearest hub. This sounds conservative. In practice, it eliminates the worst class of dispatch failures: the car that accepts a booking and then needs an emergency charge mid-route.
Every vehicle in our fleet reports state of charge, estimated range at current driving style, and proximity to a hub charger in real time. The matcher treats these as hard constraints alongside driver availability. A 90% charged Nexon EV near Cyber City beats a fully available vehicle at 18% charge near Sohna Road — every time — even if the Sohna car is technically closer to the pickup pin.
We also model charging windows into shift planning. Drivers on long-shift duty get vehicles with enough buffer for their expected trip load, and the system schedules hub returns during predicted low-demand periods. Charge management isn't a back-office problem bolted onto dispatch — it's part of the same optimisation loop.
Predictive positioning, not reactive matching
Layering predicted demand on top — airport banks, office egress at 6pm, weekend leisure — lets us pre-position charged vehicles before the requests even arrive. That pre-positioning is where most of the 38% improvement came from.
Our demand model ingests flight arrival data for IGIA, historical booking patterns by hour and postcode, and corporate calendar signals from partner accounts (Monday morning spikes, Friday evening airport runs). By 5:45pm, the system already knows which corridors will need capacity at 6:15 — and moves idle vehicles accordingly. The rider who taps 'book' at 6:12 isn't waiting for a car to cross the city. They're matched to one that was already nearby because the system saw them coming.
“The fastest pickup is the one you predicted before the rider tapped 'book'.”
The human layer
Route AI handles matching and routing. Our dispatch desk handles the exceptions — a VIP pickup with specific vehicle requests, a last-minute multi-stop itinerary, a corporate account that needs a guaranteed vehicle held for a board meeting. The algorithm optimises for the median case; humans handle the long tail. That split keeps the system fast without making it brittle.
Chauffeurs also feed signal back into the model. A road closure not yet on the map, a new security checkpoint at a corporate campus, a monsoon-season flood on the service road — these get flagged through the driver app and propagate to the fleet within minutes. Route AI isn't a static model deployed once. It's a loop that gets sharper every week we operate in NCR.
Measuring what matters
We track pickup time, but we also track pickup time variance — because a 7-minute median means nothing if the 90th percentile is 22 minutes. Corporate mobility runs on reliability, not averages. The 38% headline is the median improvement; the p90 improvement is 31%, which we're equally proud of.
There's more to come. We're training corridor-specific models for the Noida and Delhi expansions on our roadmap, integrating weather-adjusted range predictions for monsoon season, and experimenting with multi-stop batching for corporate shuttle routes. The goal is simple: make 'your car is here' feel instant — even in a city that wasn't built for instant anything.
