What wasn't working.
Manual toll booths depended on an operator identifying vehicle type and typing plate details by hand for every car. This was slow enough to create queues at peak hours.
The manual entry step was also a fraud vector. Operators could misclassify vehicles or skip entries entirely, and there was no reliable, automated record of what vehicle had actually passed through a lane.
What we actually did.
We built a computer vision system that detected vehicles, classified them by type, and read license plates automatically as they passed, removing the operator from the loop.
Once the detection and classification pipeline existed, the same underlying data, vehicle counts, types, and flow patterns, was valuable well beyond tolling. We built a second module for road authorities to analyze traffic patterns for planning and survey purposes.
Where the interesting calls were.
We chose to automate the fraud vector itself, not just the slow step. Removing manual entry entirely, rather than speeding it up, is what actually closed the fraud gap.
We reused the detection pipeline for a second use case instead of building a separate system. Traffic analytics were a natural extension of data the system already produced.
We invested in vehicle type classification, not just plate reading. Solving the broader problem, what is this vehicle and not just what does its plate say, supported both products from one model.
The numbers after we shipped.
Toll lanes ran fully unmanned, with no operator required to log vehicle or plate data.
The manual data-entry fraud path was eliminated, since there was no manual entry step left to exploit.
The same computer vision pipeline shipped a second, revenue-relevant product for traffic analytics, built on infrastructure that already existed.
Things I'd carry into the next case.
Automating a manual step often removes a fraud path as a side effect. Look for that overlap before scoping a project narrowly.
Data collected for one purpose is frequently valuable for a second. Plan the pipeline to support reuse from the start.
B2G products live or die on reliability under real-world conditions like weather, lighting, and plate wear. Test there, not just in the lab.