jobsearch v0.0.1

← cohere / Product Manager, Safety & Security

brief / art_zdq62tKMPi4

role
cohere / Product Manager, Safety & Security
model
anthropic/claude-sonnet-4.6
created
2026-05-29T18:58

Company snapshot

Cohere is an enterprise-focused AI company building frontier large language models and deployment infrastructure for developers and organizations; its flagship product North is an agentic AI platform designed for secure, compliant enterprise deployment of AI agents and automations. Cohere differentiates from OpenAI/Anthropic by targeting enterprise data privacy and on-premises/private-cloud deployment rather than consumer markets. In the last 12–24 months the company has raised significant capital (reported $500M+ Series C in 2024, though exact figures should be verified), expanded offices across Toronto, New York, San Francisco, London, and Paris, and has been actively building out its North platform with agentic capabilities. Cohere has a strong research culture with published work on command models, RAG, and safety; engineering reputation is generally positive for rigorous ML research and enterprise-grade reliability. Specific internal projects or named individuals beyond public disclosures are not confirmed.

Team stack

Based on the JD and public signals: core model stack is likely transformer-based LLMs trained on proprietary infrastructure (likely GPU clusters; framework uncertain but PyTorch is standard in the field). North product layer is likely built on microservices with REST/gRPC APIs, likely Python backends for model serving, and enterprise integrations via connectors/plugins. Safety evaluation tooling likely includes internal red-teaming harnesses, benchmark suites (likely including standard safety benchmarks such as TruthfulQA, HarmBench, or internal equivalents), and possibly RAG-specific adversarial test sets. Agentic orchestration within North likely involves tool-use frameworks, multi-step reasoning chains, and guardrail middleware layers. Specific languages/frameworks for the product layer are not confirmed from the JD; TypeScript/React for frontend is plausible given industry norms.

Likely questions (10)

areaquestionwhy
domain Walk me through how you would design an evaluation framework to detect safety regressions as Cohere ships new model versions into North. What metrics would you track and how would you gate releases? The JD explicitly calls out 'Define and drive evaluation frameworks for assessing how safety properties hold up as models and product capabilities evolve' — this is a core deliverable of the role.
domain You've built aeval, a local-first model evaluation platform. How did you decide which eval types to prioritize (factuality, safety, reasoning, etc.), and how would you adapt that thinking to enterprise LLM safety at Cohere's scale? The JD lists 'hands-on experience with LLM evaluation, safety benchmarking' as a nice-to-have; aeval is the candidate's most directly relevant evidence and interviewers will probe its depth.
system_design Describe how you would architect a guardrail layer for an agentic system like North that uses tool calls and multi-step reasoning. Where do you insert safety checks, and how do you avoid breaking legitimate workflows? The JD specifically calls out 'unique safety challenges introduced by tool use, multi-step reasoning, and autonomous execution' and asks the PM to 'coordinate development of guardrails and intervention mechanisms.'
domain Explain prompt injection in the context of a RAG-powered enterprise agent. How would you scope a research evaluation to surface this risk, and what product-level controls would you recommend? The JD lists 'prompt injection, jailbreaks, RAG poisoning' as specific threat vectors the PM must understand; this tests both technical depth and product translation ability.
behavioral Tell me about a time you had to translate highly technical research findings into a product decision for a non-technical audience. What was the finding, how did you communicate it, and what was the outcome? The JD requires 'strong written communicator who can translate complex model behavior findings for non-technical audiences' — a direct behavioral signal they will probe.
behavioral Describe a situation where safety research surfaced an unexpected finding mid-roadmap. How did you decide whether to act immediately, deprioritize other work, or accept the risk? The JD states 'safety research surfaces unexpected findings, and this role requires good judgment about what to act on and how fast' — they want evidence of judgment under ambiguity.
coding You don't need to write code, but walk me through the statistical methods you'd use to determine whether a safety regression is real or noise across model evaluation runs. What sample sizes, confidence intervals, or tests would you apply? The JD requires technical depth to 'engage credibly with safety researchers'; aeval uses bootstrap CIs, Welch's t-test, and Cohen's d — the candidate has direct experience here.
system_design How would you build a scalable safety review process for new North features as the product surface area grows rapidly? What gates, checklists, or automated checks would you put in place? The JD explicitly asks to 'build processes for scaling safety review as North's surface area grows, including how new features get assessed for safety risk before launch.'
culture Cohere describes this as 'not a traditional PM role' — you'll spend as much time reading evaluations as writing PRDs. How do you think about the balance between research engagement and product delivery, and where have you lived in that space before? The JD is explicit that this role blurs PM and research operations; they want to assess whether the candidate genuinely gravitates toward research depth or will drift toward classic PM work.
behavioral You've founded two AI companies and held Staff PM roles at Intuit and Splunk. Why safety research PM at Cohere specifically, and what draws you to the enterprise AI safety problem versus continuing to build your own products? Cohere will probe motivation fit given the candidate's entrepreneurial background; they need confidence the candidate wants to go deep on safety research, not use this as a stepping stone.

Talking points