← figma / Product Manager, AI Platform
brief / art_eYMusASpT-4
Company snapshot
Figma is a browser-native collaborative design platform used by millions of product teams worldwide to design, prototype, and hand off work. Over the last 12–24 months Figma has aggressively expanded into AI-powered product creation: Figma Make (prompt-to-production-code) and a purpose-built design agent that edits files directly on the canvas are the flagship bets. The company went through a high-profile failed Adobe acquisition (blocked by regulators, ~2023) and subsequently refocused on independent growth and AI product expansion. Figma is widely regarded as having a strong engineering culture with high craft standards, particularly around real-time collaborative systems and developer tooling. Specific internal team structures and recent personnel moves are not publicly confirmed.
Team stack
Based on the JD and public signals: TypeScript/React frontend (Figma's editor is heavily TypeScript); Go or Rust likely for performance-critical backend services (inferred from Figma's engineering blog); frontier model APIs (OpenAI, Anthropic, Gemini — based on JD reference to 'frontier labs'); MCP (Model Context Protocol) servers for agent tool integration (explicitly named in JD); vector search / embedding infrastructure for the context/search platform (likely, based on JD); eval pipeline infrastructure (likely custom, possibly wrapping open-source eval frameworks — based on JD); cloud infra on AWS or GCP (unconfirmed, likely AWS based on industry norms); CI/CD with automated quality gates (inferred from JD emphasis on 'continuous quality'); Docker/container orchestration for sandboxed code execution in Make (likely, based on JD description of code generation and preview).
Likely questions (10)
| area | question | why |
|---|---|---|
| system_design | Walk us through how you would design the eval platform for Figma's AI suite — what metrics matter, how do you run evals continuously without slowing down shipping, and how do you handle regressions? | The JD explicitly calls out 'evals and continuous quality' as a core platform area and asks for 'methodology and infrastructure that lets every Figma AI feature improve week over week.' |
| system_design | Figma Make needs to generate, preview, and publish code against real production codebases. How would you architect the sandboxed execution and preview environment to be safe, fast, and scalable across thousands of concurrent users? | JD lists 'developer systems: how code is generated, edited, previewed, and published' as a core ownership area. |
| system_design | How would you design the context platform that injects a customer's design system tokens, components, and brand guidelines into AI prompts — while ensuring enterprise customers' proprietary data never leaks across tenants? | JD explicitly names 'search & context: how Figma leverages context from inside and outside our platform while retaining the trust of enterprise customers.' |
| domain | You've benchmarked GRPO, DPO, PPO, and other RL algorithms in your workbench. How would you think about which post-training approach is right for a design-domain AI agent, and what reward signal would you use? | JD requires experience with AI agent systems and the candidate's RL Workbench evidence is directly relevant; Figma's design agent likely requires RLHF-style alignment. |
| domain | MCP (Model Context Protocol) is explicitly named in the JD as part of the agent infrastructure. How have you used or thought about MCP servers, and what are the key design decisions when exposing tools to a frontier model agent? | JD calls out 'agent infrastructure that integrates frontier models and MCP servers' as a core platform component. |
| coding | Given a stream of AI-generated code diffs from Figma Make, how would you instrument and measure code quality — correctness, style conformance, and design-system adherence — in an automated pipeline? | JD requires hands-on technical fluency ('you can read a PR') and ownership of the eval platform for code generation quality. |
| behavioral | Tell me about a time you had to make a build-vs-buy call on a core platform capability. How did you frame the decision, who did you involve, and what happened? | JD explicitly states the PM will own 'build-vs-buy calls' as part of platform strategy. |
| behavioral | Describe a situation where you were the platform PM and multiple product teams had conflicting needs from your platform. How did you prioritize and communicate your roadmap commitments? | JD emphasizes 'scaling platform capabilities across multiple product teams' and 'distill learnings into platform commitments.' |
| culture | Figma's AI platform sits at the intersection of product and infrastructure — you're building foundations other teams ship on top of. How do you stay close to end-user outcomes when your direct customers are internal engineering teams? | JD describes the role as 'intersection of product and infrastructure' and emphasizes influencing product roadmaps from a platform position. |
| behavioral | You've led 0-to-1 products as a founder and also scaled platform infrastructure at Intuit. How do you shift your operating mode between greenfield exploration and disciplined platform execution, and which mode does this role demand most? | JD asks for 'a track record of driving strategy and execution across ambiguous, fast-moving product areas' — the interviewer will probe whether the candidate can operate in a structured enterprise platform context vs. founder mode. |
Talking points
- Built aeval — a production eval platform (FastAPI, TimescaleDB, Redis, Ollama, Next.js) with 5 eval types, adversarial safety testing, bootstrap confidence intervals, Welch's t-test, and CI/CD regression gates — directly mirrors Figma's 'evals and continuous quality' platform area. Can speak to methodology, not just tooling.
- Built OpenClaw multi-agent orchestration framework with gateway protocol, subagent delegation, and session management in StreamIO — and integrated MCP SDK (Claude) for screen-capture tool exposure to AI coding assistants. Direct hands-on experience with the agent infrastructure and MCP patterns Figma is building.
- At Intuit, owned the ICE developer platform end-to-end: reduced onboarding from 2–3 weeks to minutes, scaled throughput from 6K to 50K TPS, achieved 675M+ engagements across 5 product lines. Demonstrates ability to own platform infrastructure that multiple product teams ship on top of — the exact operating model for this role.
- RL Workbench benchmarks 12 algorithms (PPO, GRPO, DPO, SimPO, etc.) across TRL, VeRL, OpenRLHF, and NeMo RL with live SSE metric streaming and GPU Docker passthrough — provides credible technical depth to participate in architecture conversations about post-training and model quality improvement pipelines.
- NeurIPS-published ML researcher (protein structure prediction, 2014) with a 2026 PyTorch rewrite spanning 413 to 8B parameters — establishes long-arc AI credibility that differentiates from PMs who are AI-adjacent but not AI-native.