19Cabs: 1115 drivers, 500 customers, 90 days from idea — and why we still had to stop and rethink AI

In the first few months, AI primarily gave us speed.

Using AI aggressively, two part-time AI engineers were able to take 19Cabs from an idea to a working ride-hailing system and into the market, without building a traditional engineering team. In roughly three months, and with no marketing, we onboarded 500+ drivers and 250+ customers in a single city — Kanpur — and processed 410+ real ride requests, a mix of completed, cancelled, and low-quality or fake rides.

These numbers are important because they came from real users, not test data. And once real usage began, the limits of “build fast” became obvious. We started seeing patterns that only appear in production: fake bookings, repeated cancellations, low-intent customers, inconsistent driver quality, and poor ride–driver matches that degraded trust on both sides of the marketplace.

At that point, we deliberately slowed the rollout. Rather than patching issues manually or adding heavy verification, we reframed these as AI engineering problems. We are now using AI to estimate ride completion probability, driver acceptance likelihood, and driver–ride compatibility, based on real signals such as location history, ratings, experience, and behavioral patterns — deciding not just who to match, but when not to dispatch at all.

The first phase of AI helped us reach the market quickly and gather real data. The next phase uses that data to fix known issues — fake rides, fake customers, and poor matching — before an official launch. Once these problems are addressed in one city, we expect a new class of challenges: scaling trust, dispatch, and reliability across multiple cities. That, again, becomes an AI problem — not just of technology, but of marketplace behavior at scale.

This talk shares what AI made possible in the first three months, what real data forced us to rethink, and how we’re using AI differently as we prepare to scale beyond a single city.