jobsearch v0.0.1

← perplexity / Product Manager, AI Capabilities

brief / art_SMlIs9NP6Nk

role
perplexity / Product Manager, AI Capabilities
model
anthropic/claude-sonnet-4.6
created
2026-09-02T19:16

Company snapshot

Perplexity AI is a conversational answer engine that combines large language model reasoning with real-time web retrieval, serving millions of daily active users across research, shopping, investing, and general curiosity use cases. In 2025–2026 the company launched 'Computer,' its agentic AI product positioning Perplexity as a platform for transforming knowledge into action rather than just surfacing information. Perplexity has raised significant venture funding (reportedly valued at several billion dollars as of 2024–2025, though exact figures should be verified) and competes directly with Google, OpenAI, and Anthropic on AI-native search and productivity. Engineering reputation is for moving extremely fast with small, high-ownership teams; the company is known for shipping frontier capabilities quickly and iterating in public. Specific internal team structures, named leaders, and recent org changes are not independently verified — claims in the JD are taken at face value.

Team stack

Based on the JD and public signals: Python backend services (likely FastAPI or similar), likely React/TypeScript on the frontend, LLM orchestration layer supporting tool use and multi-step agentic workflows (likely custom or built on top of model provider APIs from Anthropic/OpenAI/Gemini), vector retrieval and RAG infrastructure for grounding, PostgreSQL or similar for structured metadata, Redis for caching/queuing (inferred from scale requirements), Docker/Kubernetes for container orchestration (likely), and internal eval/evals frameworks for nondeterministic model assessment. The 'Skills,' 'Projects,' and 'Computer' product surfaces suggest a composable capability registry — likely a tool-use / function-calling abstraction layer. Artifact generation implies structured output pipelines. All stack details beyond Python + LLM APIs are inferred from the JD and public engineering blog signals.

Likely questions (10)

areaquestionwhy
system_design How would you design the 'Skills' registry for Computer — the data model, discovery surface, versioning, and permission model — so that both Perplexity-built and third-party skills are composable and trustworthy at scale? The JD explicitly owns 'Skills that teach Computer how to work' and asks for fluency in extensibility, composability, and ecosystem growth — this is the core platform design question for the role.
system_design Walk me through how you'd architect a 'Projects' feature that maintains persistent context, knowledge, and ongoing work across agentic sessions — including how you'd handle context window limits, staleness, and multi-user collaboration. The JD lists Projects as a first-class capability owned by this team; the question probes PM-level systems thinking on state management and collaboration in agentic contexts.
domain Agentic systems are nondeterministic. How do you build an evaluation framework that gives you enough signal to make confident product decisions about whether a new skill or agent behavior is ready to ship? The JD explicitly calls out 'work with research to evaluate and steer nondeterministic models' and 'experience with evaluations' — directly maps to the candidate's aeval platform work.
domain What is your mental model for 'composable AI capabilities'? Give a concrete example of a capability primitive you've shipped and explain how you designed it to be reusable across multiple surfaces or workflows. The JD's core thesis is 'reusable, composable AI capabilities' — interviewers will probe whether the candidate has genuine product intuition here vs. buzzword fluency.
behavioral Tell me about a time you had to make a difficult product decision under significant uncertainty — especially one involving a nondeterministic or AI-driven system where you couldn't fully predict outcomes. The JD explicitly states 'conviction to make difficult product decisions in the face of uncertainty' — this is a stated hiring signal.
behavioral Describe a 0-to-1 product you shipped end-to-end. What did you get right, what did you get wrong, and what would you do differently? The JD emphasizes ownership, initiative, and small agile teams; the candidate has multiple 0-to-1 products (Vantage, Fintellect, StreamIO) and will be expected to speak to the full arc.
coding You need to build a capability-adoption flywheel: users discover a skill, use it, rate it, and that signal improves ranking and recommendation. Walk me through the data model and the metrics you'd instrument from day one. The JD calls out 'levers for capability discovery, adoption, reuse, sharing, and ecosystem growth' and 'experience building data-driven flywheels for iterative improvement.'
culture Perplexity ships very fast with very small teams. How do you decide what NOT to build, and how do you maintain product quality and coherence when you're moving that quickly? The JD stresses 'small, agile team' and 'initiative and desire for ownership' — interviewers will probe prioritization discipline and quality bar under speed constraints.
domain How do you make complex AI agent behavior understandable and controllable to end users — what interaction models, permission patterns, or feedback mechanisms have you found most effective? The JD explicitly lists 'make complex AI behavior understandable and controllable through clear interaction models, thoughtful defaults, permissions, and well-designed feedback' as a core responsibility.
behavioral Give me an example of driving cross-functional alignment when a capability spanned multiple teams or product surfaces. How did you resolve conflicting priorities or ownership ambiguity? The JD calls out 'drive alignment when capabilities span Computer, Search, Projects, artifacts, connectors, or multiple teams' — a known coordination challenge at Perplexity's current scale.

Talking points