← adobe / Principal Product Manager
brief / art_fnjvzFmmOv0
Company snapshot
Adobe is a ~$20B revenue enterprise software company best known for Creative Cloud, Acrobat, and its Experience Cloud suite (RTCDP, AJO, CJA, AEM). Over the last 12–24 months Adobe has made aggressive moves into generative AI with Firefly (image/video generation), GenStudio for content supply chains, and the integration of AI assistants across Experience Platform — positioning AI-native workflows as the core growth vector for its CX Enterprise business. The hiring team sits inside Adobe Experience Cloud, building the agentic runtime and control plane that grounds LLMs in Adobe's first-party data products. Adobe's engineering reputation is strong in media processing and data pipelines; its platform/infra reputation in the agentic layer is newer and still being established (based on JD signals and public announcements, not confirmed internal sources).
Team stack
Based on the JD and Adobe's public Experience Platform architecture: runtime likely Python-based agent harness (likely LangGraph or custom loop, not confirmed); tool protocol layer built around MCP and A2A specs; skills/tool registry backed by Adobe's existing entitlements and IMS auth infrastructure; grounding data from RTCDP (real-time customer profiles), AJO (journey orchestration events), and CJA (analytics); likely cloud-native on AWS (Adobe's primary cloud post-Magento/Marketo consolidation); evaluation infra likely custom, possibly integrating with Adobe's existing telemetry stack; developer surface likely CLI + SDK + DevPortal pattern (consistent with enterprise platform norms); LLM providers likely multi-model (Claude, possibly GPT-4, internal fine-tunes) — all inferences based on JD language and Adobe public disclosures.
Likely questions (10)
| area | question | why |
|---|---|---|
| system_design | Walk us through how you would design the agent loop contract for Adobe's CX Enterprise harness — specifically, how do you define the boundary between the runtime core and domain-specific skills, and what does the tool protocol interface look like at that seam? | The JD explicitly calls out owning 'the agent loop and the contracts around it — tool protocols (MCP, A2A), skills, knowledge grounding.' This is the central platform design question for the role. |
| system_design | How would you design the override and feedback loop for agents running in production against enterprise customer data? What signals feed back into agent quality, and how do you prevent the feedback loop from being gamed or drifting? | The JD states 'the override and feedback loop is the moat; you'll design it' — this is explicitly called out as the defensible differentiator and will be probed deeply. |
| domain | You've used MCP in your own projects (StreamIO MCP server). What are the current failure modes of MCP as an enterprise tool protocol — schema versioning, auth delegation, context window pressure — and where would you diverge from the spec for an enterprise control plane? | The JD requires fluency in MCP and asks for 'informed opinions about what they get right and wrong.' Your StreamIO MCP server work makes this directly testable. |
| domain | In your RL Workbench you benchmarked GRPO/DPO across TRL, VeRL, OpenRLHF, and NeMo RL. How does that experience inform how you'd design agent quality benchmarks that predict customer outcomes — not vanity evals — for an enterprise agentic platform? | The JD calls out 'agent quality benchmarks that actually predict customer outcomes — not vanity evals' and asks for partnership with Applied Science. Your RL Workbench and aeval platform are directly relevant. |
| coding | Here is a simplified skill manifest / MCP server definition [expect a YAML or JSON snippet]. Walk us through what it does, what's missing for enterprise use, and what you'd change. | The JD states 'you should be able to read a skill manifest or an MCP server and know what it does' — expect a live artifact review, not just conceptual discussion. |
| behavioral | Tell me about a time you held a position with engineering leads on a platform contract decision and either successfully defended it or changed your mind. What was the technical crux and how did you reason through it? | The JD explicitly requires: 'You can hold a position with engineering leads and back it with technical reasoning, and you can change your mind when the argument is better.' |
| behavioral | At Intuit you drove a Service Language Assessment across 9 languages and presented to the CTO. How did you structure the analysis, manage stakeholder disagreement, and translate technical tradeoffs into an executive recommendation? | Adobe needs a PM who can 'translate between research, engineering, and GTM without losing precision' — the Intuit CTO-level work is the strongest signal on your resume for this. |
| domain | Adobe's control plane needs to handle auth, entitlements, and governance without reinventing what the ecosystem already does. Where do you draw the line between owning the control plane and adopting open standards — and how do you make that call in practice? | The JD directly states: 'Hold the line on control plane ownership (auth, entitlements, governance) without reinventing what the ecosystem already does well.' This is a build-vs-adopt judgment question. |
| culture | Adobe's JD says 'write tersely, strip adjectives, calibrate by audience.' Give me an example of a document you wrote — a PRD, an eng spec, a customer doc — and walk me through how you calibrated it for its audience. | The JD is unusually explicit about communication style as a hiring signal. They will probe whether your actual writing matches the terse, precise standard they've set. |
| system_design | How would you design the developer surface for two distinct personas — Agent Developers who define skills, and Agentic App Developers who compose them — so that local dev/prod parity is maintained and the abstraction doesn't leak? | The JD calls out both personas explicitly and requires owning 'docs, CLI, local dev/prod parity.' Your Intuit DevPortal and ICE Self-Service work is directly analogous. |
Talking points
- I've already built and shipped an MCP server in production — StreamIO's desktop app exposes screen capture tools to AI coding assistants via the MCP SDK, and I implemented the OpenClaw multi-agent orchestration framework with a gateway protocol, subagent delegation, and session management. I have direct, hands-on opinions about where MCP's enterprise gaps are: auth delegation, schema versioning under concurrent skill updates, and context window pressure when grounding against live data sources.
- At Intuit I owned the ICE platform end-to-end — DevPortal, GitOps config, SDK Starter Kits, and the developer onboarding surface — scaling to 675M+ engagements in FY23 and reducing onboarding from 2–3 weeks to under 24 hours for production. That's the same developer surface pattern Adobe is describing: two personas (skill authors vs. app composers), local/prod parity, and a control plane that handles entitlements without reinventing auth. I've shipped that at enterprise scale.
- My RL Workbench benchmarks 12 algorithms (PPO, GRPO, DAPO, DPO, SimPO, and more) across TRL, VeRL, OpenRLHF, and NeMo RL with standardized throughput/memory/convergence metrics, and my aeval platform implements bootstrap confidence intervals, Welch's t-test, and saturation detection for model evaluation. When Adobe says 'benchmarks that predict customer outcomes, not vanity evals,' I have a concrete technical framework for what that means and have already built the tooling.
- I've designed and shipped the full agent feedback loop in my own products: Fintellect Agents use RAG with ChromaDB, multi-provider LLM orchestration with fallback routing, and structured output validation — and AutoEval repurposed my streaming pipeline to close the evaluation loop for robot model training, cutting eval cycles from 72 hours to 4 minutes. The override-and-feedback moat the JD describes is something I've architected from scratch, not just read about.
- NeurIPS-published researcher (protein structure prediction, 2014) with a CS/engineering foundation from UC Berkeley and hands-on coding across the full stack — from hand-coded BPTT in C++ in 2004 to production TypeScript/React/Electron, FastAPI, Redis, Docker, and Swift ScreenCaptureKit integrations today. I can read a skill manifest, debug an MCP server, and write the eng spec for it — and I've taught Java, cloud computing, and data analytics at De Anza, so I calibrate technical communication across audiences by default.