Every Delivery.
On Time.
Average delivery time went from 41 minutes to 23.
Three cities, 340K deliveries a month, an ML router
that re-plans on live traffic, and ETAs that hit within
three minutes 94% of the time.
On-demand logistics & delivery platform
The old stack used static routes: drivers followed pre-set paths whatever the traffic did. Across three cities, hundreds of concurrent drivers and a 30-minute SLA, average delivery crept to 41 minutes and customer scores were sliding.
We rebuilt routing and dispatch on PostGIS spatial data with logical-replication CDC into the live optimiser, OR-Tools and Vroom + GraphHopper as the routing solver, the Mapbox Navigation SDK 3.x for turn-by-turn and Mapbox Movement for live GPS, with a Flutter driver app on top. On delivery, an on-device VLM (Gemini Nano / Apple Foundation Models) handles proof-of-delivery from the photo, and the OpenAI Realtime API powers hands-free voice for drivers. Average delivery dropped to 23 minutes. 340K orders a month, 99.4% accuracy, and NPS recovered inside weeks.
Technical Approach
-
Routing that re-plans live — OR-Tools and Vroom + GraphHopper solve the multi-stop VRP on PostGIS, fed by logical-replication CDC for live traffic, driver position, order clustering and historical patterns. Routes update as conditions change, not only at dispatch.
-
Sub-2-second dispatch — A Redis-backed dispatcher assigns an order to the right driver in under two seconds, balancing utilisation, ETA and priority. Nearby orders get batched into multi-stop runs.
-
Tracking and notifications — Mapbox Navigation SDK 3.x renders driver and customer positions live, Mapbox Movement is the live-traffic input. Twilio sends SMS and push for ETA updates, confirmations and exceptions. ETAs refresh every 30 seconds at 92% accuracy.
-
One Flutter driver app, hands-free — iOS and Android from one codebase, with offline-capable navigation so a drop in coverage doesn’t lose the route. On-device VLM (Gemini Nano / Apple Foundation Models) checks the proof-of-delivery photo on the spot; the OpenAI Realtime API powers a hands-free voice mode so drivers don’t fight the screen at speed. Shifts, earnings and history all live inside the app.
What was actually hard
Three cities, three street grids, three different rush-hour shapes. One late order is forgivable, two in a row and the customer is gone. We had to pull average delivery from 41 minutes to under 25, keep accuracy above 99%, and make ETAs honest enough that the customer trusts the next one. The routing model had to care about reality, not the map.
Project Outcome
Average delivery dropped from 41 minutes to 23 — a 44% cut that reset what customers expected. 340K deliveries a month across 3 cities at 99.4% order accuracy, and driver utilisation climbed 31% because the router stopped sending people on empty legs.
users > 340K deliveries
per month > 23 min avg delivery
(from 41) > 99.4% order
accuracy
“Average delivery dropped from 41 minutes to 23. That's not optimization — that's a different business.”
@ Carlos R.
COO — On-Demand Delivery Startup
More Case Studies
