The startup teaching city buses to file paperwork - and quietly running the largest moving traffic-camera fleet in America.
It is a Tuesday morning on First Avenue in Manhattan, and the M15 is moving. That sentence is its own small miracle. For most of the past decade the M15 has been a parable - a bus, in theory, stuck behind double-parked sedans, idling FedEx trucks, and the occasional sedan whose driver believes the bus lane is a personal driveway. Today it is moving because somewhere behind its windshield there is a small box, full of NVIDIA silicon, watching everything that gets in its way.
That box belongs to Hayden AI. The company has been pre-classified, internally and politely, as "smart-city infrastructure." That undersells it. Hayden AI is what happens when a handful of computer-vision engineers decide the most interesting AI problem in the country is not chatbots or copilots, but the slow, unglamorous question of how to make a city bus on time.
The company sits at 460 Bryant Street in San Francisco, occupying the kind of unmarked door that startups choose when the work is more interesting than the branding. About 190 people work there. They have raised roughly $246 million. They have, by their own count, more than 650 cameras now riding US transit buses, photographing law-breakers from inside a windshield while the bus does what buses are supposed to do, which is keep going.
For roughly a century American cities have built bus lanes, painted them red, and then watched them be treated as suggestions. Studies pile up. Reports get filed. Buses crawl. Riders stop being riders. Drivers, smelling a vacuum, take more lanes.
The traditional fix is a police officer with a ticket book - a fix that does not scale, sometimes carries social cost, and almost always misses the actual violation by the time the cruiser arrives. The other fix is fixed cameras, which only see the corners they were bolted to.
Hayden AI's founders looked at this and asked an obvious question with an inconvenient answer. What if the camera moved? What if it rode the bus itself?
The competitive landscape isn't another startup. It's inertia - the assumption that traffic enforcement must be human, occasional, and political. Hayden's competitors are the parts of city government that don't talk to each other, and the cynical voice that says nothing can be changed on a street where everyone is already breaking the same rule.
Hayden AI was founded in 2019 by Chris Carson, Vaibhav Ghadiok, Bo Shen, and Michael Byrne. Carson took the CEO chair; Ghadiok runs the technology. The bet they placed was small and stubborn: don't sell cities a new piece of infrastructure - sell them a new use of the infrastructure they already pay for.
That meant the camera had to be small enough to live behind a windshield without a retrofit. The inference had to happen on the bus, not in a data center. The evidence had to be admissible in front of a judge or a hearing officer who has, frankly, seen enough AI demos for one career. And the cost per citation had to be lower than the cost of a parking enforcement officer plus the cost of a delayed bus plus the cost of the collision that delayed bus might cause.
The architecture they landed on is unfussy in a way you only get from people who have already shipped things. An NVIDIA Jetson module does perception at the edge. TensorRT keeps the models fast. AWS handles what's left. The whole stack is documented in a US patent granted for the behind-the-windshield camera system - a quiet win that puts a moat around an idea most cities had not yet realized they wanted.
The product is, on one hand, a camera. On the other hand it is an opinion: that public transit is the cheapest, fastest, most carbon-honest way to move a city, and that defending its right-of-way is a problem AI is finally good enough to solve.
It is one thing to install 650 cameras. It is another for the cameras to have done anything. The data from the New York deployment - the largest in the country - is where the company's story starts to feel less like a pitch deck and more like an argument.
Source: Hayden AI program data, MTA + press disclosures · A chart that fits in a tweet, and also in a city council hearing.
Customers now include MTA in New York, WMATA in Washington DC, LA Metro, AC Transit in the East Bay, Santa Monica's Big Blue Bus, and a slowly lengthening list of pilots. In 2024 the company signed a commercial agreement with Verra Mobility, the long-established player in automated road enforcement, to scale the work across more cities. The Series C, $90 million led by TPG's Rise Fund, was less a vote of confidence than a vote on the math.
It would be easy, and slightly wrong, to call Hayden AI a traffic enforcement company. The cameras enforce. The mission is bigger. The mission, in the language the company uses internally and in the language its investors clearly bought, is to build the operating system for an autonomous traffic-management platform - and to do it in a way that pays for itself.
That last part matters. Public transit budgets are not a place where pretty AI demos go to live. Anything Hayden installs has to return value the city can write down on a line item. Citations help. So do the longer-term wins: faster buses, fewer crashes, lower emissions because more riders choose transit when transit actually works.
Putting AI cameras on hundreds of city buses is not a free move. Hayden has been deliberate about the privacy framing - on-device processing, evidence packages that capture only what's needed, citizen-facing dispute portals, an emphasis on automated rather than ad-hoc police interaction. Whether that framing holds up over time is one of the more interesting policy questions of the next decade. The company appears to know that.
You can see the tags they wear in public.
privacy-first AI smart city edge inference B2G SaaS computer vision transit safety digital twin climate
The most consequential AI deployments of the next ten years will not look like a chatbot. They will look like a slightly more honest bus lane. They will look like an intersection that responds to the bus it sees rather than the schedule it was given in 1987. They will look like fewer ambulance trips and more retirees who feel comfortable taking the bus alone at 8pm.
Hayden AI's bet is that this kind of AI - quiet, specific, deployable, accountable to a city council - is also the kind of AI that scales. Every additional bus is another sensor. Every additional sensor is more data for the digital twin. Every digital twin nudges urban planning into a tighter feedback loop with reality.
That is a long way of saying: the company is small now and will not be small for long.
Return to the Tuesday morning on First Avenue. The bus is moving. The cameras behind its windshield have logged a half-dozen violations and the riders inside do not know it. They are looking at their phones. They are getting to work. They are not, for the first morning in a while, getting off and walking because the bus has surrendered to the traffic.
That, finally, is the whole pitch. Hayden AI is not trying to make a city feel more high-tech. It is trying to make a city feel like a city: predictable, fair, in motion. The AI is just the boring part underneath. If they get it right, you will stop hearing about them. The bus will simply arrive.