The case was closed. Detectives had worked a case in which a man was wrongly convicted, and they had been able to exonerate him. Then they put a question to Peregrine, as the company's Ben Rudolph tells it on the Training Data podcast: "Hey, can you reproduce this result with an agent?"
Peregrine sells software to public-safety agencies, connecting the records those agencies already hold so staff can search and analyze them. An agent, in the sense Rudolph uses the word, is not a chatbot answering a single question. It is software that plans its own steps, works through material with tools, and keeps going for a long stretch before reporting back. The cold case agent was the first one Peregrine built for the outcomes its customers care about, rather than for its own engineers.
Boxes this big, with CDs attached
Rudolph describes what a high-profile investigation physically looks like. An agency uploads 200 to 300 gigabytes of material: video, audio, images and, as he puts it, "a ton of PDF files." A single detective is tasked with going through all of it, which on its own takes a really, really long time. He says he has been in police departments where the records are still on paper, in boxes he sizes with his hands, with CDs attached to them. "It's really an outrageous amount of data."
That backlog is the problem the agent was pointed at. Peregrine built it together with the customer, using the evidence and data the department had already assembled for that case.
Reproducing what the detectives found
The agent would run for 30 minutes, 60 minutes, to start to glean insights. Eventually, Rudolph says, it reached the place where it could reproduce the results the detectives had gotten to.
The test therefore ran over a closed file whose answer was already known, and success meant arriving at findings human investigators had reached first. The investigation and the exoneration were the detectives' work and the legal system's outcome.
Peregrine's own product page for its investigations assistant describes the mechanics it offers: ingesting case files, transcribing and translating recordings, turning handwritten notes and documents into searchable text, and pulling out timelines, people, places and networks. Supported material includes call-detail records, financial documents and social-media warrant returns. Findings are linked back to the evidence they came from, down to a spreadsheet row or a table inside a PDF, and the page says the assistant cross-references case material with the agency records a particular user is authorized to see, returning citations and the steps behind a conclusion so a person can check it. The company describes the assistant as available in early access with selected agencies. The page is product documentation; it does not speak to the specific cases Rudolph describes, neither of which he names.
A few phone pings in 300 gigabytes
The agent is now in use in a few departments across the United States. Rudolph describes recent work in a county in Wisconsin on a similar type of case, again with about 300 gigabytes of data, where investigators were able to identify and place the suspect not only at the scene of the crime but at the location where the body was found.
What carried it, he says, was a small thing: a few cell call detail records, the trace of a phone pinging a network, scattered among an enormous amount of other material. Placing a suspect is a step in an investigation; Rudolph does not describe what followed in court.
For Rudolph, that example points at the part of the job he thinks is worth taking off people's hands. He talks about upleveling and unburdening public servants from the administrative weight of the work: watching hours and hours and hours of video, listening to hours and hours of audio.
"The quiet professionals"
His co-founder Nick Noone broke in to say why the story matters inside the company, and why Peregrine mostly does not tell it. The way the business maintains trust, he says, is by not taking credit and not shouting from the rooftops about the awesomeness of what happened. Scooping up a customer's work and trumpeting the company's own skill is, in his view, one of the fastest ways to break trust with these organizations. "Being the quiet professionals in a context like this and empowering the customer is why we have access to the next problem."
The host pushed back: he wished they would talk about it more. Public distrust in AI is so high at the moment, he said, and this is a wonderful story.