How Crosstown Finds Stories Hidden in Municipal Data

In 2018, Gabriel Kahn, a journalism professor at USC Annenberg, walked into a meeting with computer scientists at his university. They told him about a contract with the Los Angeles Metropolitan Transportation Authority. Sensors embedded in freeways capture the moment wheels pass over them, generating roughly 400 million data points every hour. From that, you can calculate the speed of every mile of LA freeways at every hour of the day, going back years.

Kahn asked what questions LA Metro (the Los Angeles Metropolitan Transportation Authority) had asked about the data. "So far they haven't asked us any questions," one computer scientist said.

"I thought, boom, there's a news product," Kahn told me.

From this observation Crosstown originated - now a software platform that ingests municipal data and turns it into hyperlocal journalism in several metropolitan regions in the U.S. 

In Los Angeles, one person writes a story template. Then the system generates 114 unique newsletter versions, one for each L.A. neighborhood. Open rates exceed 90 percent. A year ago, Crosstown spun out from USC as an LLC and now offers its platform as a white-label service to newsrooms including the Philadelphia Inquirer, Nextdoor Chicago and WRAL in Raleigh, North Carolina.

Cities publish enormous amounts of data about themselves: crime, arrests, building permits, complaint calls, parking tickets, property assessments, restaurant inspections, etc. In Los Angeles alone, Crosstown collects 15 different datasets, updated daily, weekly, or monthly. The problem, Kahn explained, is that this data is "publicly available but not publicly accessible. You need a PhD in data science to make it accessible."

Crosstown's engineering team, led by Luciano Nocera, a computer scientist and Associate Director at USC's Integrated Media Systems Center, built a dashboard that lets journalists query the data in seconds. "We lowered the barrier to entry," Kahn said. "I could get undergraduate students to come up with a hunch, type it into the dashboard, and get an answer. If the hunch was wrong, they wasted 30 seconds. If it was right, they had a story."

Every row of city data includes geographic coordinates. In L.A, Crosstown maps those coordinates onto 114 neighborhood shape files. When a journalist writes a newsletter, the system automatically populates each version with the relevant local numbers, charts, and context. The platform, Nocera confirmed, was built entirely from scratch with no third-party vendors for core functionality.

During our conversation, Kahn demonstrated the dashboard by querying new restaurant openings in Los Angeles. The chart showed that 2024 had the highest number of new restaurant openings on record. "The L.A. Times wrote a whole bunch of stories about how terrible it is for the restaurant business," Kahn said. "They actually don't know that more and more restaurants have opened. It's true that many have closed, but many more have opened."

This is the core value proposition: Data surfaces stories that human reporters miss if they are working without systematic access.

Crime data offers another example. Kahn pulled up 20 years of burglary rates in Chicago, where Crosstown also operates. The long-term trend showed a dramatic decline, punctuated by a recent uptick. "News has traditionally done crime by incident," he said. "Here's a terrible incident, let's roll the news trucks. That only creates panic. When you use data to contextualize the situation, you see most of these situations are actually getting better."

Crosstown's foundation is machine learning with rules-based automation: Structured data goes in, templated text comes out. That architecture cannot hallucinate because no facts enter the text except what exists in the verified dataset. But Generative AI is now expanding what the system can do.

Nocera explained to me that outlier detection uses transformer-based models to find trends and anomalies that would be invisible to a human scanning the data manually. An early prototype called "Detective" scoured Los Angeles crime data and flagged oddities. One finding: 27 cars stolen from the same address at the same hour. Kahn investigated and discovered that a gang had held up a parking garage attendant, taken all the keys, and driven off with the cars. "It was like a heist movie," he said. "But it popped up in the data."

The next step, Kahn said, is having AI ask thousands of questions of the data every day and surface a prioritized list of story tips. "Rather than getting an inscrutable list of prompts, you'd get something like, 'Car theft is down, but in these three neighborhoods, it saw a 20 percent or more increase.'"

Hallucination risk remains a concern. Nocera's approach: "You do repeated requests so you can check your results and confront that against data you know are true." The journalist remains in control and is not a passive consumer of AI output. "The opportunity," Nocera said, "is to turn journalists into editors who can use tools offline to create stories that go around all these problems."

According to the Reuters Institute's 2025 report on generative AI and news, only 12 percent of people say they are comfortable with news produced entirely by AI. Crosstown's newsletters are auto-generated from templates. Does that matter to readers?

Kahn argues that the distinction lies in sourcing. "If this were total AI slop, it would probably say midnight is the most dangerous moment of the day," he said. Police data often defaults to 00:00 when the actual time of a crime is unknown. Pure automation would misread that quirk. "You have to understand the idiosyncrasies of the data. The biggest mistake is trusting the data blindly."

Entirely AI-produced crime news would also likely insert teasers promising to reveal specific specific names and addresses - clickbait like “Click here to find out which of your neighbors has been stealing your package deliveries.” But Nocera points out that all the public datasets Crosstown uses are anonymized. Crime data shows block-level locations, never exact addresses, and includes no names of suspects or victims.

For Crosstown, the goal is “not to chase the latest trends that big tech has produced," Kahn said. Instead, he wants to recreate something more timeless - the feeling of finishing a newspaper and knowing you are informed. Kahn's goal is to recreate something he remembers from the International Herald Tribune. "It was 24 pages. You got through it. You're like, okay, I'm done." He wants that feeling of completion at the neighborhood level.

"I don't need time spent on site. I don't like any of these other metrics. The only metric that I want is: does that person at the end of this experience feel that they have gained knowledge?"