First, which Peregrine. The name is crowded: a UK aerospace and defense contractor hiring embedded C++ developers, Peregrine Semiconductor (now pSemi, the RF chip company), the long-dead Peregrine Systems enterprise-software firm, and the falcon. This guide is none of those. It’s Peregrine Technologies, the San Francisco company that sells a data platform to police departments, fire, 911 centers, and other public-safety agencies.
What Peregrine actually does explains the whole interview. A police department’s data is scattered across a dozen systems that don’t talk to each other: computer-aided dispatch from one vendor, a records management system from another, body-camera footage in a third, license-plate-reader feeds, permits, 911 logs. Peregrine ingests all of it and makes it searchable in real time, without the agency handing over ownership of the data. That’s the Palantir Foundry and Gotham playbook pointed at local government, which is no accident: co-founder and CEO Nick Noone ran Palantir’s US Special Operations business before leaving to start Peregrine, and co-founder Ben Rudolph came from the UN Refugee Agency and Dimagi. The two embedded inside the San Pablo, California police department to build the first version, and the forward-deployed model stuck. The company says eight of the eleven US host cities for the 2026 FIFA World Cup ran on Peregrine over the summer. One tell in the hiring data: Peregrine’s H-1B filing record includes six “Deployment Strategist” positions, the Palantir name for the engineer who sits with the customer and makes the platform fit their reality.
For peers to calibrate against, Peregrine sits with the government and defense-adjacent companies, not the consumer-AI labs. The Palantir interview guide is the obvious lineage read: product, culture, and half the leadership come from there. The Anduril interview guide is the closest gov-tech peer on hiring bar and mission framing. The broader AI-native company interview guides hub collects the rest. Be realistic about level before you apply: the postings skew toward engineers who’ve shipped real systems, and the open SF roles as of late 2026 include lead-level positions like an Applied AI Engineering Lead, so a new grad is swimming upstream unless a posting explicitly welcomes one. “Shipped real systems” means a few years owning something in production that other engineers depended on, beyond closing tickets, so a solid mid-level engineer (three to five years) is in range. Core software-engineer and senior-engineer titles sit below the lead level and appear in the company’s hiring filings, even if the applied-AI opening on the public board was pitched at lead.
The round structure you should expect
No public Glassdoor record or first-hand candidate write-up exists yet, which is normal for a company this young. What follows is reasoned from the postings, the forward-deployed culture, and how comparable data-platform companies hire. Treat it as a grounded prediction, not a leaked rubric, and confirm specifics with your recruiter.
| Round (reasoned from Peregrine’s postings and Palantir-lineage culture, not a confirmed Peregrine loop) | Who runs it | What it likely screens for |
|---|---|---|
| Recruiter screen (30 min) | Recruiter | Motivation, role/track fit, comp expectations, work authorization, comfort with the public-safety mission |
| Technical / coding (60 min) | Engineer | Practical coding: parsing messy inputs, data transforms, API and data-structure work. More “build the thing” than abstract puzzle |
| Data / systems design (60 min) | Senior engineer | Designing the ingestion-and-search layer: schema mapping across disparate sources, freshness, access control, audit |
| Applied-AI or domain round (60 min) | AI/ML or product lead | Entity resolution, search relevance, and where an LLM helps versus where it’s a liability on police data |
| Values / mission (45 min) | Founder or exec | Judgment, customer empathy for agencies, and a real view on the civil-liberties tradeoffs of police data |
Five rounds is typical for a well-funded data-platform company hiring fast but protective of culture; expect it to compress if you’re senior and referred. For a Deployment Strategist or forward-deployed role rather than a core-platform one, expect less algorithmic coding and more of a case: here’s a tangled customer data situation, model it and tell me what you’d ship first.
The coding round maps to messy real data, not LeetCode
Peregrine’s daily work is taking half-broken government data exports and turning them into something queryable, so a one-hour coding round is likelier to resemble that than a graph-theory brainteaser. Plausible shapes: parse a CSV or fixed-width dump where the schema drifts between files and reconcile it against a canonical model; dedupe records that refer to the same person or vehicle across two systems that spell and format everything differently; join two feeds on a timestamp when one is local time, one is UTC, and neither is clean. You can still get standard data-structure reasoning, so keep the common coding patterns and your Big-O analysis sharp, but the differentiator is judgment about dirty data, not reciting algorithms. SQL fluency helps too; much of this work is joins, window functions, and dedup queries over imperfect tables.
System design: build the search layer over data you don’t own
A realistic prompt: design the system that ingests a department’s CAD, records-management, body-cam metadata, and license-plate-reader feeds and lets an analyst search across all of it in near real time. Treat it as a distributed systems design problem with two constraints a generic “design search” answer misses.
The first is schema heterogeneity. Every agency’s RMS models an “incident” differently, so the hard part isn’t the search index, it’s the mapping layer that normalizes dozens of source schemas into one common model and stays maintainable as sources change. Talk about an ingestion pipeline per source, a canonical entity model, and indexing tuned for the lookups analysts run. The second constraint is that the data isn’t yours. The agency retains ownership, which means hard tenant isolation, per-agency access boundaries, and a complete audit trail of who looked at what. Say the words access control and audit log early, because on criminal-justice data they’re what keeps the company compliant and trusted. Expect follow-ups on data freshness (how stale can a feed be before it’s dangerous for an officer relying on it), partitioning by agency, and what happens when a source goes down mid-ingest. Name the failure modes out loud: a public-safety platform that silently serves stale or wrong data is worse than one that admits it doesn’t know.
The applied-AI round: entity resolution and where not to trust a model
The open Applied AI Engineering Lead role signals this is a real track, not a bolt-on. The center of gravity is less “train a big model” and more applied: entity resolution (deciding two records are the same person despite different spellings and typos), search relevance over a mix of structured and unstructured data, and natural-language querying so a non-technical analyst can ask a question in English. Realistic prompts: how would you decide two incident records describe the same event; how do you rank results when a query spans structured fields and free-text narratives; where would you put an LLM in this product, and where would you refuse to.
That last one carries weight. A strong candidate says plainly that a language model is fine for summarizing a long narrative or turning a plain-English query into a structured filter, but is a liability as the sole arbiter of whether a person matches a record, because a hallucinated match is a real person wrongly implicated. Keeping a human in the loop on consequential decisions is the mature answer. For handling AI tooling during the interview itself, our guide to interviewing in the AI era is worth a read.
The values round, and the part people underprepare
A company that sells surveillance-adjacent tooling to police departments knows it’s in a live civil-liberties debate, and the values round will probe whether you’ve thought about it. Walking in with no view reads as naive, not neutral. You don’t have to be a true believer; you have to show you understand the tension between a tool that genuinely helps solve crimes and one that enables overreach, and that you’ve thought it through. Because the founders built v1 embedded in a real department, customer empathy matters here too. Bring real stories with specifics; the STAR method keeps them tight.
Two logistics items to raise on the first call, not at the offer stage. Vendors that touch criminal-justice information systems typically require staff to pass a CJIS fingerprint-based background check; expect it and ask how it applies to your role. And on work authorization, Peregrine does sponsor H-1B (its public DOL record spans several engineering titles, with new Labor Condition Applications filed as recently as 2025) but had no PERM green-card filings on record, so if you need long-term sponsorship, ask how far they’ll go rather than inferring it. If you’re not a US citizen, also ask whether your work-authorization status affects CJIS access, because that varies by agency and contract rather than being a blanket rule.
Compensation: what’s grounded, and what isn’t
The careers-board index doesn’t surface salary bands, but California pay-transparency law requires the individual SF postings to carry a range, so the posting itself is where to look first; anyone quoting an exact Peregrine range without it is estimating. The most verifiable public figures are the company’s own H-1B filings, which list real base salaries by title: a senior software engineer at $180,960, a software engineer at $161,637, and a network engineer at $195,000, with Deployment Strategist roles in the $120,000 to roughly $159,000 band. Those are single filed positions, not published scales. Levels.fyi shows a median total comp near $200,000 for a Peregrine software engineer but no sample size, and a Recruiting From Scratch estimate puts the all-role base median around $210,000 ($170,000 to $238,000) across dozens of postings, which spans every job, not an engineering band.
The real lever at a company that went from $2.5B in March 2025 to $6.8B fifteen months later is equity, and the gap between your 409A strike price and the preferred-share price is where the value sits. Ask for the latest 409A, the preferred price per share, and the fully diluted count, then model the grant with the total comp calculator and treat that, not the base, as the number to push on in any salary negotiation.
What separates a strong Peregrine candidate from a strong generic one is narrow and testable. You’ve wrestled real, inconsistent operational data into something usable and can talk about what broke. You treat access control and audit as first-class design concerns, not afterthoughts. And you’ve decided you can build this particular product with your eyes open. Bring those three and the rest of the loop is ordinary engineering interviewing.
Practice the behavioral round:
