Nick Noone and Ben Rudolph were turned down more than two dozen times before the San Pablo police department let them into the building, on 26 February 2018.
Public-safety technology companies have grown by selling agencies more places to record things — cameras, sensors, forms for officers to fill in. Peregrine sells none of that. It joins up records a city already holds, leaves ownership with the agency, and charges for the joining.
"For us, we almost want to decentralize that: to get to a world where we are not, in any shape or form, in the business of bringing more data to the customer that they don't already have."
Noone ran Palantir's unit for US Special Operations Command and deployed into intelligence operations in the Middle East; Rudolph built data systems for the UN Refugee Agency on the Sudanese and Colombian borders. Peregrine now works with police, fire, emergency medical services, emergency management and health services.
I listened to the full interview so you can skip it. 52 minutes of audio, 23 minutes of reading.
Here are the 17 takeaways that matter.
👤 Guests: Nick Noone, co-founder and chief executive of Peregrine, who ran Palantir's unit for US Special Operations Command before starting the company; and Ben Rudolph, co-founder, who built data systems for the UN Refugee Agency and at the global health technology company Dimagi
🎙️ Host: Sonya Huang, a partner on Sequoia Capital's growth team and a host of the firm's Training Data podcast
📰 Published: 1 September 2026 on YouTube (Sequoia Capital)
🔴 YouTube | 🟣 Apple Podcasts | 🔗 Show notes | ⏱️ 52 min | ✅ Time saved: 29 min
Key Takeaways
Peregrine grows by connecting the records a city already holds, not by collecting new ones Noone called it the inversion of a model in which vendors sell more collection systems to customers whose data they already hold
A cold case agent reproduced an exoneration that detectives had reached by hand In a Wisconsin county, phone records buried in 300 gigabytes of evidence placed a suspect where the body was found
Agents write about 90% of the code that connects a customer's data to the platform They run for hours, split into subagents and report back to an orchestrating agent
Most of the work in a public-safety deployment happens before anyone asks the system a question Rudolph said the engineering effort goes into making the data answerable rather than into the model
The founders were turned down more than two dozen times before a police department took them in They got in by asking a commander to teach them rather than by selling him anything
Peregrine will not take a company-wide position on facial recognition Noone said most American public safety agencies and their communities choose not to use it
Forward deployed engineers are treated as research and development, not as the cost of a sale A workaround one engineer built, updating a record through its comments, told the company to build editing into the product
Serving small cities meant getting the cost of a deployment below a million dollars a year
1. Own the customer's outcome
Huang opened by asking what forward deployed engineering actually means, given Noone spent his Palantir years doing it.
Noone said he had no idea what the term meant before he joined Palantir, and learned it by watching colleagues leave the office for customer sites and come back with stories
The first of his two rules is that the engineer takes on the customer's problem as their own. "First: we psychologically go in and own, or co-own, the problem. We talk internally about getting all the way to the outcome, with and on behalf of our customers." He put a speed target on it: "Technology, and all the skills and ways of delivering our tech, is part of the answer — but the real answer is just getting to the outcome at all costs, ideally three to five times faster than any other person or team could do it."
The second rule is that the result belongs to the customer. "And second: recognizing that at the end of the day, it's the customer's win. It's not our win."
He described the appeal in plain terms — "It felt to me like the adventure of a lifetime" — and said the reward is the effect on the institution and on the people across the table
2. Moving fast reads as ego
Noone said the thing Silicon Valley misses is how its confidence lands inside an institution. Smart engineers from selective schools arrive at a place like the Los Angeles Police Department and try to absorb 30 years of institutional context quickly — not only the data but how the organization works, how decisions are made, how teams interoperate
The customer does not believe that is possible, and reads the attempt as arrogance. "And unfortunately, I think that shows up as ego to a lot of organizations, where people have dedicated their entire lives to a particular role."
The alternative is harder than it sounds. "Letting go of our intelligence, letting go of our own skills and abilities, trying to suspend our ego and really get into the customer's context — it's an easy thing to say, but it's such a high-empathy, patient way of working."
3. Downstream of data problems
Huang turned to Rudolph, who spent the same years doing refugee work in Africa and India and arrived at a similar view of what technology is for.
He turned down a job at Airbnb to work for the UN Refugee Agency, and deployed to the Sudanese border and the Colombian border, where he says he learned where technology helps and where it does not
The agency's problem was not effort but disconnection. "UNHCR is this deeply disconnected organization. It does phenomenal work, but it's deeply disconnected." Its data sat in spreadsheets and was hard to make sense of
That produced the thesis the company runs on. "That's where I started to form this thesis that a lot of these problems are downstream of data problems, where you can have a lot of impact."
After the refugee agency he joined Dimagi, a company that builds last-mile healthcare software in under-resourced places, and worked with the Indian government on applications to help people in rural India stay on their tuberculosis drugs — technology that then spread across the country
4. Safety is the base layer
The two met as gymnastics teammates and, on a vacation about ten years after graduating, agreed to build something together — Noone flying between Baghdad and other posts in the Middle East, Rudolph flying into Africa
They chose cities because that is where people spend their lives, and then worked backwards to what a city needs first
Safety came out as the foundation, in two forms. "And what we came to is that at the bottom of the pyramid, when it comes to how to make cities awesome, is the idea of safety: that people need objective safety and also need to feel safe."
They wanted two things that sound opposed. Building the backbone infrastructure of a modern city could preserve individual privacy while also bringing people together, and Noone said that is what he finds beautiful about cities: "if we're staring at the same information, we kind of want the same thing"
5. Two dozen no's before yes
The count of rejections is a blur, but large. "You know, it's such a blur at this point, but it's definitely over two dozen, in my opinion."
The founding date is the day they got through the door, not the day they incorporated. "We started the company on February 26th of 2018 — we'd incorporated a month or so earlier, but February 26th is the day the San Pablo, Northern California police department permitted us to step into the building with our badges." They had desk space and access to information
They found their way in by researching practitioners rather than buyers. An article online about Operation Red Reach, a cross-jurisdictional narcotics investigation in Contra Costa County, led them to Brian Bubar, a young commander who had risen fast inside a rank-bound organization
The approach was a request to be taught, not a pitch. "We just called him and said: Brian, we don't know that much. We may have some utility, but I don't know." They asked to come in, ask questions and maybe build something over time
Noone credits the framing for the access. "We led with that level of trust, and that was what finally got us access."
6. Building through defund
Huang asked what it felt like to start a policing company as the "defund the police" movement gathered force.
Noone described holding two things at once. "We had very close friends, people we'd graduated with and built things with in the past, who were standing outside the Women's Center in San Francisco protesting the police. And we were driving a Honda Accord across the Bay Bridge out of San Francisco every day to work with gang homicide investigators. We were holding those two truths at the same time."
His argument is that proximity changes the judgement. "To me, it's not enough to look at a very difficult situation from the outside. The privilege is jumping into the pool and swimming." The deeper they went, he said, the more gray zones there were
Recruiting was the concrete cost. At the time the company worked only with police departments, and Noone said engineers would not even respond to a recruiting call
The business is broader now — police, fire, emergency management and health services — with what he called a through-line of delivering safety and prosperity to cities
He now describes the product as deliberately apolitical. "I think we've seen the pendulum shift, and we're trying to deliver solutions that are apolitical: what people, communities, and the institutions that serve them truly want and need."
7. The inverse of Flock, Axon
Huang asked how the business differs structurally from data-collection companies such as Flock and Axon.
Noone described the incumbent model as growth through more collection. Hardware or software is installed inside the customer, the business is about the input, and "Historically, those companies have grown because once they have information that belongs to the customer inside their system, plus a distribution advantage, they can sell more stuff — and the additional stuff tends to be more collection systems."
Peregrine's model runs the other way. "Peregrine is the inversion of that entire model." The company joins disparate information into a governed layer that sits on top of systems the agency already has
The product is precision, not volume. He described the business as building solutions that drive greater precision in how a person interacts with their own data, rather than bringing that person more data
Concentration is the thing he says the public fears, and the company's answer is to disperse it. "For us, we almost want to decentralize that: to get to a world where we are not, in any shape or form, in the business of bringing more data to the customer that they don't already have."
The bottleneck is usability, not supply. "The fundamental problem is that they can't utilize the data they have — or utilize it in a way that's secure and high-trust for the communities they serve."
8. The customer owns the data
Rudolph said permission controls, data governance and sovereignty have been the design constraint since the first deployment, because without them these institutions cannot buy at all
Ownership is stated flatly. "We deeply believe in the idea that each customer, each institution, owns their own data." The organization serves the community, so the data is the community's and the organization's — "It's not Peregrine's data."
The controls are built to allow narrow sharing rather than none. Agencies can roll the system out internally and share selected pieces when they need to, with controls specific enough that they do not overshare
The comparison is to what agencies do without a system like this. "Without these types of solutions, folks will put a bunch of data in the back of a car and drive it across the city — and there's actually less control there."
Huang put the reconciliation in her own words, saying the data being locked down is how the tension between public safety and a surveillance state gets resolved: local intelligence, local data, everything locked down
9. From nice search to why
Asked what customers actually do with the system, Rudolph said the answer changes as departments get used to it.
The first thing a department gets is better lookup. "When you initially deploy this type of AI to these organizations, you often just get what I'd call nice search." Instead of scanning rows for an address, a user sees everything that happened at that address in one formatted view
The interesting work starts once the floor rises across the whole department, and questions that were previously impossible become answerable
A Florida county asked why it was doing so many rescues. "We're working with a county in Florida, and the other month they had to do over 100 water rescues. And they asked: why? We've never had to do this many water rescues before." Noone explained the term for listeners: "A water rescue is — when there's a flood, rescuers need to get on a boat and literally go out and rescue someone from their car, which might be stalled in the middle of the flood, or from their home."
The answer came out of several passes of agent research, not a single query. The weather patterns had occurred before, but never on three consecutive days; three days of them cut sand channels, the channels produced rip currents, and the rip currents produced the rescues. Incident records, 911 calls and weather data were all in the same place for the system to work across
Rudolph said the finding could not have been anticipated from headquarters. "And we just couldn't have made this up. If we were sitting in our headquarters trying to pontificate on what emergency responders in a hurricane needed, there's zero chance we would have figured that out."
The second example is a keyword problem that keywords cannot solve. A detective investigating a threat against a synagogue wanted to know whether there had been other antisemitic threats against two synagogues in the area. "When you think about that question as a detective — what are the keywords you search for to find that information? It's actually very difficult for keywords to find it." Searching on meaning rather than words, Rudolph said, surfaced a pattern of threats against those synagogues
10. Preparation is the product
Almost all of the effort sits ahead of the question. "Certainly 95% of the work is what happens before the user types in the question: all the preparation you do to get to a place where the AI can answer accurately." Rudolph called getting data AI-ready a hard problem and said most of the company's engineering effort goes there
The second shift is what cheap software generation lets a deployment team do. "What happens when the cost of generating software is virtually zero?" His answer is a platform that hands the deployment team security controls, governance controls and application programming interfaces, and then lets them write whatever the customer needs
Not every customer wants an agent. "What we're seeing is that these deployment strategists — you know, not every customer needs an agent; maybe they need a hurricane simulator — are able to build these things really well, with AI."
Customers can build in the platform, but mostly they build with Peregrine's team. Rudolph said the company enables customer-built applications, and that some customers do it, but the common pattern is a partnership in which the deployment team supplies the technical work
He said that division is already moving. Asked whether it would change over the coming years, Rudolph said he sees it changing now, and that embracing it takes confidence — "let's innovate, and may the best method win"
Noone said the technology layer is deliberately replaceable. The domain expertise and the outcome stay fixed while the technology changes: charts, then agents, then whatever follows, all of it useless without accurate underlying data. "So I believe the core strategy of Peregrine has remained very unchanged in this new age of AI."
11. Agents write the pipelines
Huang asked whether the company is doing anything with long-horizon or background agents. Rudolph answered with what the platform is.
Half the engineering organization works on the layer that ingests a customer's existing data. "About 50% of our engineers spend most of their time working on the data platform that enables our deployment team to integrate data the customer already owns."
The scale of that work so far. "At this point we've integrated tens of thousands of datasets across all these customers."
The integrations are now mostly machine-written. "At this point — we use a version of Python notebooks to do a lot of our integrations — about 90% of that is written by agents, with the oversight of our deployment team." The company built an evaluation set that scores those agents deterministically for completeness and correctness on integrations
These are long jobs, not prompts. "And these agents run for hours." They analyze the databases, work out the ontology, and split into subagents that report back to an orchestrating agent
Huang compared it to using coding agents for codebase migrations — long-running, unglamorous work — and Rudolph agreed, adding that the reason it works is that the result can be checked: "I love this problem because it's verifiable, and that makes the problem a lot easier — this is why coding agents are, in a lot of ways, a lot easier."
12. The cold case agent
The scale of a high-profile investigation is the reason to automate it. "To give you some context on what some of these very high-profile investigations look and feel like: you're uploading 200 to 300 gigabytes of data." Video, audio, images and a large number of PDF files, all of which a detective has to go through "I've been in police departments where they have paper records of these cases in boxes that are this big, with CDs attached. It's an outrageous amount of data."
The first agent was commissioned by a customer as a test against a known answer. The department had worked a case in which a man was wrongly convicted and was later exonerated. "They said: hey, can you reproduce this result with an agent?"
The build ran against that case's own evidence. "So we took all the evidence and data they had on that case and started to work on an agent that would run for 30, 60 minutes and start to glean insights." Eventually it reproduced what the detectives had found, and the agent is now in use in several departments across the US
The Wisconsin case is the one Rudolph says shows the value. "We were recently working in a county in Wisconsin — again, a similar type of case, 300 gigabytes of data — and they were able to identify and place the suspect not just at the scene of the crime, but where the body was found." "And this was from a few cell call detail records — the ping of a phone — scattered among a lot of data."
He framed the gain as removing administrative work from public servants, rather than as replacing their judgement — no more watching hours of video and listening to hours of audio to find the few records that matter
13. The anti-network effect
Noone said the company deliberately does not publicize results like the Wisconsin case. "The way we maintain trust is by not taking credit — not shouting from the rooftops about the awesomeness of what happened here." Claiming the win is one of the fastest ways to break trust with an institution "Being the quiet professionals in a context like this, and empowering the customer, is why we get access to the next problem."
Huang pushed back on that policy. "I wish you would talk about it more, though. We're in this moment where public distrust in AI is so high, and this is a wonderful story."
Rudolph's answer is that the data landscape has enormous network effects and that Peregrine's job is to resist them. A company hacking its way to something useful could pull in open-source data sitting in a gray zone under laws, regulations, ordinances or a department's own procedures — and he said no organization needs to break those rules to introduce AI
He named the position. "It's almost the anti-network-effect proposition: how do you preserve the sanctity of the data, the ownership of the data, and the ways of working for every individual agency and organization — and then build the connective tissue?" "The interoperability is going to happen." What decides whether it happens well, he said, are the unglamorous parts: data governance and permissioning logic in the context of AI
Huang gave the fear its name. "A lot of this data used to just exist — and I think people now fear what happens when it's all swept up into a central panopticon. That's deeply un-American."
Rudolph described the discipline that follows as restraint about growth. "I find that our business is a big practice in letting go." The questions are how to build transparent solutions, how to make that legible to customers and their constituents, and how not to grow too fast
14. Follow customer, follow law
Huang asked about morally loaded technology choices, using facial recognition as the example, and about the company's North Star for hard decisions.
Noone rejected the idea of a vendor setting policy for an industry. "I think the idea of a Silicon Valley company imposing a general-purpose decision on an industry is fundamentally wrong." That applies to whether a technology is used and to how long data is retained
The company's role is to lay out the considerations and then wait. He described helping a customer understand its own context, bringing to light considerations it may not know about, and — with patience — helping usher in the right way of deploying a technology
On facial recognition he reported what agencies decide, rather than what Peregrine prefers. "On facial recognition specifically: most public safety agencies in America, and their communities, choose not to implement it." Some take a strong position, for others it is a matter of preference
A red line drawn by the vendor is the thing he will not do. "But creating a technological red line without understanding the texture and context is not a boundary we think we can assert on top of our customer."
Huang summarized it as "follow the customer, follow the law," and Noone added a third step: make the limits visible. He said the company tries to create clarity about what the left and right limits are and what the institutional norms may be
Agencies learn from each other more than from vendors. "It's astounding how cities call each other. They call each other for advice." Helping them get the best information faster, he said, changes decisions — including decisions about which technologies to use
15. Selling to the underdogs
Huang framed the contrast with Palantir, which she said "doesn't take anything less than eight-figure contracts," and asked whether people told them the economics would not work.
Noone's answer was that the depth is what makes it scale, not what stops it. He said it takes confidence to believe that something built well solves problems that have never been solved before, and that it will scale later
The stack had to be built end to end. "The idea of going deep to build a vertically integrated tech stack — everything from network access to the permissioning, the governance, the ETL, the pipeline, the ontology, the UX, the APIs — configuring all of that and creating a system that's open, interoperable, and can check those boxes... the audacity, I think, is real." He added that each component had to be first-class rather than a checked box In plainer terms: the company built its own connections into agency networks, its own permission and governance controls, its own pipelines for extracting and loading data, its own model of how a city's records relate to each other, and its own interface
The market had never been able to buy this class of software. "We were truly thinking about how to deliver these types of technology solutions to an end market that has never been able to use them before — let alone at the price point we were able to deliver them at."
The method looks unscalable from outside. "That might seem extremely manual, extremely laborious, extremely unscalable." What they found, he said, is that doing it produced a platform that did something different for these organizations and also scaled
16. Sending people into a cave
Huang said the deployment strategists she has met are unusually good and asked how the company selects and trains them.
Noone said he and Rudolph took tests to align on character qualities, and that the pursuit of excellence was one of them; the culture is a feedback culture and he acknowledged it can be intense
The model is trust in the individual, described through an analogy. "We use this analogy of a dark cave. We'll send a person into a cave, and they have a tool belt, and hopefully the tools are pretty good." There may be a rope with a colleague at the entrance, so someone can pull them out "But truly, we are sending people into these zones and saying: maintain your integrity, maintain your principles, here are the tools — go do good."
The belief underneath it is that innovation comes from the edge of the company, and that engineering, product and design are fed by what the forward deployed team finds across the country and now in several countries
R&D, not a cost center
Huang said most companies she talks to treat forward deployed teams as a cost center. Noone put them in a different budget. "Definitely R&D — definitely R&D and growth."
Expansion inside an account is run as a series of pilots. After landing, the team keeps sprinting on additional use cases — "how we land and then expand in a customer base is fundamentally about leading from the front through our forward deployed engineers"
The company does not design products in a room. Noone described coding on the drive to San Pablo for a specific detective's request, and Rudolph said the request was real and urgent, from a named officer
The innovation lab
A workaround told the company what to build. "For the longest time, Peregrine had no ability to edit specific fields and properties. So one forward deployed engineer, to get around that, made an integration based on the comments people would write on these objects: the comment would be read by Peregrine and would then update a property." The company saw it and built editing The instruction that produces such workarounds is blunt: the engineer's job is to hit the objective, and Noone's phrase for the rest is "To hell with the technology."
The second category is work that never comes back into the product, because it is too specific to one customer — what Noone calls the innovation lab "I was just looking at one the other day where someone had built a hurricane simulator in Peregrine, and I was like: how did you do that? I didn't know you could do that." "There was another deployment strategist who built this thing that integrated data from all these different sources — 911 call times, budgets — and you could place a fire department in the city, and it would give you an estimate of how many people that would impact." Huang noted that the person who built it had been at the company for a month
17. Below $1M a year
Huang's closing question put the company ten years out, as the institutional memory layer for 10,000 cities, and asked how they think about that power.
Noone answered with the operating model rather than the ambition. "The end game of 10,000 cities requires an operating model that is fundamentally about infrastructure, actually." The company supplies infrastructure so agencies can do what they want, and what they are required to do, with their own data
Uniformity is not the goal. "Every city, every jurisdiction, I deeply believe, is unique, and I think the idea of preserving that uniqueness is actually beautiful."
The economics are what made the next customer possible. "The marginal cost of doing what we do — to drop it below a million bucks a year — is radical."
Larger companies have tried this market and failed on price. He described a lineage of failed business units inside major organizations that could not deliver tailored solutions at a price state, county and city agencies could afford, and said cutting that price by an order or orders of magnitude is what earned Peregrine the right to try
He put the power with the customer rather than the vendor. "And who are we to think that we hold any power over that? It's the institution that has the power."
Bonus Insights
Huang opened by describing Peregrine as a company most listeners have not heard of that is "almost certainly keeping you and your loved ones safe," and framed the episode as being about AI that protects cities while rejecting the surveillance state
The first ride-along was with Oakland's police department, overnight, in a suit and tie. The founders said most Silicon Valley visitors doing ride-alongs then were reporters or community organizations rather than engineers, so they got what one of them called "the white-glove service" — and it showed them how hard these institutions are to build trust with
Noone described himself as an experience-based, tactile learner who builds understanding through action, which he said leads him to question assumptions without tipping into cynicism
Rudolph turned a question back on the host, asking what she thought about customers building their own applications, before answering it himself
Noone's bottom line is that a public-safety software company can be trusted with a city's data only if it never owns that data, and that the deep, manual way of deploying it is what eventually made the price low enough for a small city to pay.
Products, Companies & Tools Mentioned
Peregrine (The company: a governed layer that joins the records a city already holds, now used by police, fire, emergency medical services, emergency management and health services)
Palantir (Where Noone learned forward deployed engineering, and the comparison Huang used for a company that she said takes nothing smaller than eight-figure contracts)
Flock Safety and Axon (Huang's examples of data-collection companies; Noone said their model grows by selling more collection systems, and that Peregrine inverts it)
San Pablo Police Department (The first customer, which let the founders in on 26 February 2018 after more than two dozen rejections elsewhere)
Oakland Police Department and the Los Angeles Police Department (The first overnight ride-along, and Noone's example of an institution with 30 years of context a newcomer cannot absorb quickly)
UN Refugee Agency (Where Rudolph deployed to the Sudanese and Colombian borders; he called it a deeply disconnected organization doing phenomenal work)
Dimagi (The last-mile healthcare software company where Rudolph worked with the Indian government on tuberculosis drug adherence)
Airbnb (The startup offer Rudolph turned down to take the refugee agency job)
Python notebooks (The format Peregrine's integrations are written in, about 90% of them now by agents)
Books & Resources Mentioned
Operation Red Reach (The online article the founders found about a cross-jurisdictional, gang-affiliated narcotics investigation in Contra Costa County, which led them to the commander who let them into San Pablo)
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: