Imagine a city where drones fly overhead, not for delivery or surveillance, but to find and report piles of trash before they fester into blight.
Illegal dumping is a global problem. The United States alone saw 1.5 million tons of waste dumped illegally in 2011, and the number has been rising every year since. For cities like San Francisco, the cost runs into the billions—wasted fuel, overwhelmed public works crews, and neighborhoods trapped under the weight of neglect.
Enter Brian Johnson, a technologist with a knack for solving gritty urban problems, and a man we’re proud to call a Tavus employee. His project marries drones, computer vision, and civic reporting pipelines to make street cleaning not just more efficient, but radically proactive.
Related: Tavus ethics and environment impact statement.
From proposal to pilot
At one point, Brian submitted a research proposal to the National Science Foundation (NSF). The pitch was straightforward but ambitious: build the world’s largest aerial dataset of illegal dumpsites, use AI to detect them, and route cleaning crews with precision.
While NSF funding didn’t land (yet)—partly because his models weren’t transformer-based at the time—he didn’t stop.
On his own dime, he flew 700+ drone missions, acquired over 100,000 aerial images, labeled more than 12,000 dumpsite annotations, and reported 6,000+ sites to San Francisco’s 311 system.
The result? Crews cleaned every pile he flagged, from sofas on sidewalks to sprawling debris fields in the Bayview. In just two months, daily dumpsites on his beat dropped from 180 to 10.
The tech stack behind the cleanup
Brian didn’t just pilot drones—he built an entire detection pipeline:
- Inverse-street and sidewalk masking: separating the street from the clutter for more accurate inference.
- Vehicle masking: removing cars from the dataset so the system could focus on actual waste.
- Trash detection models: trained on his self-labeled dataset, experimenting with ways to balance shape bias and texture bias in computer vision
At the time, his experiments with “stylized training” (think: applying Van Gogh or Monet filters to dumpsite images to force models to see shapes rather than just textures) were ahead of the curve. Initial results showed a 24% jump in model performance.
Scaling challenges and shifting rules
One of the biggest barriers wasn’t technical—it was regulatory. For years, flying drones over cities required special FAA waivers. That slowed down adoption, with smaller cities telling him: “We’ll buy the software, but we can’t fly the drones.”
Fast forward to now: FAA rules are evolving to prepare for drone deliveries at scale. That regulatory shift reopens the door for municipalities who already own drones and have certified pilots. Brian believes this is where the project’s SaaS potential lives—selling detection software to cities that can handle their own flights.
Why this matters for the future of cities
The stakes are high. Illegal dumping isn’t just an eyesore—it costs U.S. cities nearly a billion dollars a year in cleanup.
It also disproportionately burdens vulnerable communities, where trash piles linger longer and compound into cycles of blight and crime.
By automating detection, cities can:
- Cut fuel costs and emissions from patrol trucks.
- Free up crews for other critical infrastructure work.
- Deliver more equitable cleaning, so neglected neighborhoods get the same response as wealthier areas.
As Brian's early pilots showed, consistent cleanup also deters future dumping. Cleaner streets lead to less dumping—a virtuous cycle cities can’t afford to ignore.
The bigger picture
When we think about technology tackling climate or waste, we often picture big national initiatives. Brian's work reminds us innovation can be hyper-local, born from one person deciding to code, fly, and annotate until the system changes.
He sums up his motivation best in his own NSF pitch: “We can make Robert [a San Francisco public works driver] more effective, and impact the whole world.”
From Wall-E warnings to real-world wins, Brian Johnson’s project shows what happens when tech meets the environment—not in theory, but block by block, dump by dump.
Why this matters at Tavus
It’s not a coincidence that Tavus is comprised of passionate, driven people.
We’re builders who believe technology should be used to solve greater problems, protecting what’s meaningful and pushing boundaries forward.
We cover AI news, trends, ethics, etc., because it’s the operating environment in which we (and our customers) live, and we’re shaping it together. We’re investing in a future where AI works safely, equitably, and at scale.
That’s the kind of future Tavus exists to create.
But before the future arrives, feel free to check out Brian’s project at Aerbits.AI, or his recent work on Sparrow at Tavus.
