SewerAI.

A camera crawls a pipe in Houston. Software in California is already taking notes.

It is barely sunrise on a Tuesday and a CCTV crawler is sliding through 16 inches of vitrified clay under a Houston cul-de-sac. The technician above ground is on his second coffee. The crawler is on its fourth manhole of the morning. Once, this footage would have shipped on a hard drive to an inspector who would scrub through it frame by frame, frowning, looking for cracks, roots, offsets, the slow geometry of failure. That inspector still exists. He is just no longer alone.

SewerAI is the small Walnut Creek company that put a second pair of eyes on every frame of that video, and then a third, and then a fourth. The eyes belong to a neural network trained on hundreds of thousands of hours of pipe interiors. The notes are NASSCO PACP codes - the alphabet soup that municipal engineers use to grade the health of their underground. The notes are now generated in minutes. They used to take days.

This is what the company looks like in 2026: a quiet AI shop wired into 2,000 cities, 30,000 miles of pipe, and a workflow that most of America never thinks about until the street collapses.

America buried a trillion-dollar asset and lost the manual.

There are roughly 1.3 million miles of public sewer pipe in the United States. Most of it is old. Some of it is very old. Almost none of it has been inspected in any systematic, queryable, digital way. The standard practice, until recently, was to send a camera down, save the file, and hope you would remember what you saw the next time the basement flooded.

The dirty truth of the sewer industry was that the bottleneck was never the camera. Cameras got better, smaller, cheaper. The bottleneck was the human being asked to watch thousands of hours of dim footage and grade every defect, on a scale of one to five, without falling asleep. NASSCO certification helps. Coffee helps. Neither scales.

By the late 2010s the backlog had a shape. Cities had more video than they had eyeballs to watch it. Engineering firms were quoting six-month turnarounds. Contractors were swimming in tape. And the rehab budgets that depended on all that coding sat frozen, waiting for someone to finish reading.

Two industry veterans walked into computer vision.

In 2019, Matt Rosenthal and Billy Gilmartin started SewerAI with a thesis that sounds obvious now and sounded faintly ridiculous then. Computer vision had become embarrassingly good at recognizing things in images. Sewer defects, it turned out, were just things in images. The trick was that nobody who could build the model wanted to spend their afternoons looking at sewers, and nobody who looked at sewers for a living was going to start training neural networks.

Rosenthal and Gilmartin had spent careers in the pipeline inspection world. They had the data, the relationships, and a stubbornness about the fact that the industry had been promised better software for two decades and still got desktop applications that crashed at lunch.

Their bet was double-sided. First, an AI engine - AutoCode - that could read NASSCO PACP defects directly from inspection video. Second, a cloud platform - Pioneer - that would replace the cracked filing cabinet of how the rest of the work happened: uploads, submittals, reviews, capital planning, the slow back-and-forth between contractor and engineer and city.

The people who signed up first

One platform. Several pieces. Almost no hard drives.

SewerAI's stack is a model of restraint. Most infrastructure software is sold by the bolt-on; SewerAI sells one thing that branches. Pioneer is the home. AutoCode is the brain. Everything else hangs from those two.

The numbers, with the asterisks left in.

What follows are the figures SewerAI cites publicly. They are large and they are growing. Take the AutoCode accuracy number with the usual industry skepticism - "99%+" depends on the defect class, the camera, the lighting, and the surveyor who labeled the ground truth. But the rest of the table is doing the heavy lifting on its own.

"Superhuman capabilities" for the people doing supersized work.

The official mission statement reads: equip utilities, engineers, and contractors with superhuman capabilities to solve supersized problems, collaboratively. It is, as mission statements go, more specific than most. The collaboration piece is the giveaway - SewerAI is not interested in replacing the inspector or the engineer. The platform is built around the assumption that there is still a human in the chair, double-checking, signing off, owning the outcome.

The AI is the second pair of eyes. The cloud is the filing cabinet. The humans remain the experts. This is a quieter pitch than most AI companies make, and it is probably why utilities, who have heard a lot of pitches, are signing.

The future of cities runs through the parts of cities you never see.

The federal Bipartisan Infrastructure Law put real money into water and wastewater for the first time in a generation. That money has to be spent, and spending it well requires knowing which pipes are actually breaking. The cities that figure out their underground get to plan. The cities that do not are going to keep repairing the same intersection.

SewerAI is not selling visionary AI. It is selling a workflow that converts old video into actionable rehab plans. The novelty is that it works. The interesting part is what happens when it works for everyone.

It is back to that Houston cul-de-sac. The crawler reaches the end of the run. The technician packs up. By the time he is back at the depot, the footage is uploaded, the defects are coded, the report is drafted, the priority score is calculated, and somewhere in a city engineer's inbox there is a note that says, gently, that this segment has moved up the list.

The street has not collapsed. The basement has not flooded. The 6 a.m. emergency call did not happen. This is the unglamorous shape of progress in infrastructure: things did not break, and somebody, somewhere, knew it before anyone else.

Channels.