After Two Dozen Police Departments Said No, They Built an Anti-Network-Effect AI Company

A county in Wisconsin handed Peregrine a cold-case file of about 300 gigabytes: video, audio, images, and a mountain of PDFs. Ben Rudolph has been in departments where the paper still lives in boxes that size, CDs taped to the side. A detective reading all of that by hand is not a realistic plan. Peregrine's cold-case agent ran for 30 to 60 minutes, pulled a handful of cell-phone call-detail records—pings, scattered through the pile—and placed the suspect both at the crime scene and where the body was found.

That agent is the kind of work Sequoia partner Sonya Huang puts at the center of this Training Data episode. One of the most important stories in AI, she says, is from a company you may not have heard of, that is almost certainly keeping you and your loved ones safe. Peregrine sells AI into cities, counties, and states for public safety, and it is doing it while rejecting the surveillance-state script. The description notes the platform now reaches law enforcement, EMS, fire and rescue, and other services in more than 400 cities and communities.

The cofounders arrived from opposite edges of the same problem. Nick Noone ran Palantir's SOCOM unit and deployed into high-stakes intelligence operations in the Middle East, with Baghdad a regular commute. Ben Rudolph graduated with an offer from a then-small Airbnb and took the UN Refugee Agency instead, deploying to the Sudanese and Colombian borders. Together they are trying to answer a question a lot of AI founders are stuck on: what does the path into the most sensitive, closed institutions actually look like?

Huang opens on forward-deployed engineering, and what Silicon Valley gets wrong about it. At Palantir, Noone had no idea the job existed until he watched people leave the office, disappear into customer environments, and come back with stories. It felt like the adventure of a lifetime: go into a customer's context, own the problem, and still take pride in the fact that the problem belongs to the customer. Two truths from that work became the company. First, psychologically co-own the problem and get all the way to the outcome, with and on behalf of the customer, ideally three to five times faster than any other team. Technology is part of the answer; getting to the outcome at all costs is the real one. Second, the win is the customer's. The magic is the impact on the institution sitting across the table.

What Silicon Valley gets wrong is more interesting. When really smart people from top schools walk into a place like the LAPD and try to quickly understand 30 years of institutional context—not just the data, but how the organization works, how decisions get made, how teams interoperate—the idea that moving fast is even possible reads as counterintuitive to the customer. Unfortunately, Noone says, it shows up as ego in organizations where people have dedicated their entire lives to a role. Letting go of your intelligence, suspending your ego, and getting into the customer's context is easy to say. It is a high-empathy, patient way of working.

Rudolph's path was the geographic opposite. Obsessed with how technology makes impact, he chose UNHCR over Airbnb and was deeply humbled on the Sudanese and Colombian borders. He came away thinking the agency does phenomenal work and is also deeply disconnected: some data, often in spreadsheets, very hard to make sense of. That is where he formed the thesis that a lot of these problems are downstream of data problems. He left for Dimagi, a small company building last-mile healthcare in under-resourced places, and worked with the Indian government on an app to help rural patients adhere to tuberculosis drugs. The technology spread throughout the country. Impact, he decided, has to happen in the communities where people actually live.

In 2016 the two old friends went on vacation with a bunch of gymnastics teammates, ten years after graduation, and asked whether they should build something together. Noone was flying back and forth to Baghdad; Rudolph was commuting God-knows-how-many flights into Africa. They looked at themselves and said the city is where people live, thrive, and spend the majority of their lives. Deconstructing work with American cities, they put safety at the bottom of the pyramid: people need objective safety and also need to feel safe, and when that stability is there, amazing things become possible. They wanted backbone infrastructure for the modern city that would preserve individuality and privacy while still bringing people together—because if everyone is staring at the same information, they kind of want the same thing.

Then came the cold calls. Noone cannot remember the exact count, but it was definitely over two dozen nos. The break was an article about Brian Bubar, a young commander catapulted up the ranks of a paramilitary organization that usually takes years to reach leadership. Bubar had been a covert operator on Operation Red Reach, one of Northern California's most iconic cross-jurisdictional gang-narcotics investigations, work that changed the state of safety in Contra Costa County. They called him and said: Brian, we do not know that much. We may have some utility. Can we come in, ask questions, learn from you, maybe build something over time, and see what happens? They led with that level of trust. On February 26, 2018, San Pablo PD let them in with badges, desk space, and access to information. That is the day they count as the start of the company.

The research phase had already included an overnight ride-along with Oakland PD, in a suit and tie. Most Silicon Valley people doing ride-alongs then were reporters or community groups, not engineers, so they got the white-glove service. It was eye-opening how hard it is to penetrate and build trust with these organizations. The lesson of that period, Noone says, was that they just had to take their time.

The early years collided with COVID, protests, and "defund the police." Noone is a tactile, experience-based learner with a personal affinity for people who feel misunderstood. He describes holding two truths at once: close friends standing outside the Women's Center in San Francisco protesting the police, and the two of them driving a Honda Accord across the Bay Bridge every day to work with gang homicide investigators. Looking at a difficult situation from the outside is not enough; the privilege is jumping in the pool. Engineers would not even return recruiting calls. The business is more diverse now—police, fire, emergency management, health services—with a through-line of safety and prosperity for cities, delivered as apolitical solutions that communities and institutions actually want. Done well, he says, that is not controversial at all.

Huang asks the structural question: how is this different from data-collection companies like Flock or Axon? Most of the firms that came before, Noone says, built on input—hardware or software installed in a customer base, sensors collecting passively or humans typing things in. Once customer information sits in their system, plus a distribution advantage, they sell more stuff, and the additional stuff tends to be more collection systems. Peregrine is the inversion of that entire model. It does not bring the customer more data. It joins disparate information the customer already has, and sits a more secure, properly governed solution on top of the pre-existing systems so people can work with more precision.

The world, he says, is very concerned about the amalgamation of data in some central authoritarian body. Peregrine wants almost the opposite: decentralization, never in the business of bringing the customer data they do not already have. The fundamental problem is that they cannot use what they have, or cannot use it in a way that is secure and high-trust for the communities they serve. Rudolph adds that permission controls, data governance, and sovereignty have been product ingredients from day one, because they are necessary to deploy into high-stakes institutions. Trust has to run both ways: with the community those public servants serve, and with the public servants themselves.

Each institution, Rudolph says, owns its own data. It is the community's data and the organization's data, not Peregrine's. The company provides controls so a department can roll the system out securely and share select pieces when it needs to. Without that, people put a bunch of data in the back of a car and drive it across the city, which is actually less control. Huang's summary is the balance Peregrine is trying to hold between public safety and the surveillance state: local intelligence, local data, everything locked down.

Facial recognition is where the philosophy gets tested. Noone thinks a Silicon Valley company imposing a general-purpose decision on an industry is fundamentally wrong. Most public-safety agencies in America, and their communities, choose not to implement facial recognition; some take a strong stand, some treat it as preference. Creating a technological red line without understanding texture and context is not a boundary Peregrine thinks it can assert. The job is to help a customer see the context, bring considerations to light, and patiently usher in the right way of deploying. Cities call each other for advice. Helping them get the best possible information so they can make their own decision, including which technologies to use, goes a long way.

When AI first lands in these organizations, Rudolph says, you often just get nicer search: look up an address without scanning rows, and see everything that happened there, formatted cleanly. As users get used to the product, the floor rises across the department, and analysis that used to be impossible starts to show up. A Florida county had to do more than 100 water rescues in a month—boats pulling people from stalled cars and flooded homes—and asked why. After a few iterations of deep research, the pattern came out: those weather patterns had happened before, but never for three consecutive days. Three days cuts sand channels, sand channels are perfect for rip currents, and rip currents put people in the water in danger. Incident data, 911 calls, and weather were all in the system; the agency could overlay terrain maps and act. If the founders had sat at headquarters pontificating about hurricane response, Noone says, there is zero chance they would have figured that out.

A second case: a detective investigating a threat against a synagogue wanted to know whether there had been other antisemitic threats against two synagogues in the area. Keyword search is a bad tool for that question, because you do not know what to type. Semantic understanding of the information the detective already had access to pulled out a pattern of threats. Huang hears embeddings and reasoning models; Rudolph's sharper point is that 95 percent of the work happens before the user types the question. Getting data AI-ready so the system can answer accurately and cite accurately is the hard problem, and a lot of Peregrine's engineering effort sits there.

Long-horizon agents are running on both sides of the house. About half the engineers work on the data platform that lets the deployment team integrate data the customer already owns—tens of thousands of datasets across customers. They built an agent, with a deterministic eval set for completeness and correctness, that lets deployment strategists integrate that data agentically. Integrations still live in a version of Python notebooks; about 90 percent of that code is now written by agents, with human oversight. The agents run for hours: they analyze databases, understand the ontology, split into subagents, and report back to an orchestrator. Rudolph loves the problem because it is verifiable, which is why coding agents are, in a lot of ways, easier.

The first operational agent was the cold-case agent, built with a customer who had already exonerated a wrongly convicted man by hand and asked whether an agent could reproduce the result. They fed in all the evidence, let the agent run 30 to 60 minutes, and eventually matched the detectives' conclusions. It is now in a few departments across the US, including that Wisconsin county. The point, Rudolph says, is to unburden public servants from hours of video and audio. The wildest field hacks are also a product radar. For a long time Peregrine had no way to edit fields, so a forward-deployed engineer turned comments on objects into an integration that wrote properties back; the team saw it and shipped editing. Another strategist built a hurricane simulator; another, a month into the job, wired 911 call times and budgets so you could drop a fire station on a map and estimate how many people it would reach. Noone's coaching line is: your job is to hit the objective, to hell with the technology. The unscalable, hacky things are the best signaling network.

After the Wisconsin story, Huang almost begs them to talk about it more: public distrust in AI is so high, and this is a wonderful story. Noone's answer is the operating system. The way they maintain trust is by not taking credit, not shouting from the rooftops about the awesomeness of what happened. Scooping up the work and trumpeting their skill is one of the fastest ways to break trust with these organizations. Being the quiet professionals, and empowering the customer, is why they get access to the next problem.

Huang titles the thesis: Peregrine, the anti-network-effects business. The agencies they serve already have unbelievably strong network effects. A company hacking its way to usefulness could hit open source and pull in data that sits in a gray zone of laws, ordinances, or department procedure. Peregrine's rule is the opposite: do not break the rules to introduce AI. Protect the data and the organizations that, by structure, already have those network effects. Preserve the sanctity of the data, the ownership, and each agency's ways of working—then build the connective tissue. Interoperability will happen. The mundane work of data governance and permissioning logic in an AI context is, to Noone, unbelievably important. Growing too fast is a practice they are trying to let go of.

The other half of the model is price. Palantir is notorious for not taking less than eight-figure contracts. Peregrine went after underdogs that had never been able to use this kind of technology, at a price they could afford. Going deep on a vertically integrated stack—network access, permissioning, governance, ETL, pipelines, ontology, UX, APIs—each component first-class, looked extremely manual and unscalable. What they found is that owning the outcome produced a platform that could show up, do something different, and also scale. Dropping the marginal cost of a customer below a million dollars a year is, Noone says, radical. A lineage of failed business units at major companies tried this market and died on the combination of tailored delivery and a price cities could pay. An order-of-magnitude cheaper is what earned Peregrine the right to keep going.

Huang has met the deployment strategists and calls them extraordinary. Noone's training image is a dark cave: you send a person in with a tool belt and a rope, a friend at the entrance if they yell, and you tell them to maintain integrity and go do good. Doing that at scale requires institutional trust and a belief that innovation happens at the farthest fringes. Forward-deployed is not a cost center at Peregrine; it is R&D and growth, the feed for engineering, product, and design across the country and now multiple countries. Ten years out, Huang sketches Peregrine as the institutional memory layer for 10,000 cities. Noone is careful. The endgame is an infrastructure operating model that lets each unique city do with its data as it likes and as it is required to do. The path there is integrity in the ways of working, not the words. And who are they to hold power over that? "It's the institution that has the power."

The hour is a public-safety story, but the operating rules travel. If you sell a complicated product, Peregrine's standard is to get to the outcome with the customer, ideally three to five times faster than any other team, and to be explicit about how far that responsibility runs. When you do a retrospective, the first question is whose win it was. Put the credit on the customer's name, and the next problem tends to find you. Huang wanted them to shout about Wisconsin; Noone would rather stay invited.

The unscalable field hacks are not a mess to clean up later. They are a weekly product-signal network: comments used as an editor, a hurricane simulator, a one-month employee's fire-station model. Build the primitives those workarounds are screaming for, and leave the one-off innovation-lab pieces in the customer's world. On the AI product itself, permissioning, governance, and audit are day-one work in a high-stakes industry, not a patch before launch. Without them, the fallback is a trunk of files and a lot less control.

On contested technology choices, Peregrine's north star is not a Silicon Valley red line. Help the customer see the texture, the left and right limits, the way peer cities already advise each other, and then let the institution decide—facial recognition included. Eight years in, the proof they offer is inverted: give the data back, give the credit away, drop the price into a range small agencies can pay, and the business still works. That is a To-B rule, not only a public-safety one. The institution holds the power. The vendor's job is to make each city more awesome in its own way.