5 Questions from the "Saurabh PM" Interviews (2025 Season)

A peek into how I hire product managers for GenAI, data, and systems-heavy roles. Note: This article was crafted by a human, with a little help from ChatGPT-5 for editing and clarity.

I recently published a long-form post on LinkedIn (read it here) inviting product managers to join my team.

That post shared a massive question bank — but in this follow-up, I want to zoom in on the actual questions I used in interviews. These aren’t generic prompts. They were designed to test:

• First-principles thinking
• Comfort with ambiguity
• Ability to reason across UX, data, and infrastructure layers
• Core values, conviction and individuality (top priority)

In each interview, I’d pick 2–3 of these at random. Then, I’d flip the script: the candidate got to interview me. Why? Because great PMs aren’t just “answering machines” — they ask, shape, and challenge. By giving them space to lead, I get a better sense of their confidence and judgment. It also levels the playing field and makes the experience fun and collaborative.

Above all, I’m looking for your core values, conviction and individuality. Every PM is different. There are no universally right answers in product - just thoughtful ones backed by conviction and user empathy.

Hopefully this will help you prepare for other interviews.

Q1. What is your opinion about the role of the semantic layer in the modern data stack?

This question checks how you navigate buzzwords: not whether you memorize them.

In the GenAI world, terms like “semantic layer” are thrown around casually. A thoughtful PM should pause and ask: “What do you mean by semantic?” Then ground it in user reality: How does this solve a real-world pain?

Most data today is poorly documented. Metadata is a mess. A semantic layer helps attach meaning to data, making it queryable, discoverable, and explainable. The problem? Historically, building glossaries or mappings was time-consuming. But LLMs may finally make it scalable.

Your job is to de-jargon the stack and ask: Does this make life better for my users?

Never get intimidated by technology. It’s a means to an end — and the end is always a real human pain point.

Q2. A high-profile customer says your GenAI product “hallucinates.” What are you going to do?

You can start by asking: What does “hallucinate” mean in this context?

Here’s what I’d tell you: the customer asked for “sales by state,” and the system returned numbers for a state they don’t even operate in. Now, they bring this up in every business review.

You want to meet the customer and understand the issue? Sorry — that’s been done. Ten times. I’ve been in those meetings too, and walked out clueless. They don’t want apologies. They want clarity.

Most candidates say they’ll talk to engineering. Good instinct, but still wrong. This isn’t just an engineering bug. It’s a trust failure.

What I want to see is your ability to trace where and how the hallucination entered the system, and whether we can expose that logic transparently to the user.

For example: was this issue in schema mapping? Prompt generation? Data coverage gaps? Or was the LLM fabricating content?

This question is a user experience problem disguised as an architecture or customer success problem. It tests whether you can move beyond fixing the system and start fixing trust.

Q3. You're the PM for Zappos. What product feature would improve the shoe-buying experience online?

This is a real-world problem: shoe fit.

Returns are high because sizing is unpredictable. Customers waste time; retailers lose money. Personalization and discovery matter too, but the biggest pain point is often fit confidence.

Some candidates suggest AR/VR or camera-based sizing. Not bad, but they quickly stumble over logistics: where’s the camera placed? Is this scalable?

The best answers? Turn this into a machine learning problem:

• Zappos has rich return data. What if we model user journeys:

Now layer in:

• Gender
• Brand
• Shoe type
• Fit preference

You can train models to generate fit confidence scores per user or SKU. And over time, you can improve user experience without needing new hardware.

Q4. What is your biggest failure as a PM?

This is where I filter out resume polishers.

Every seasoned PM has scars - some painful, some embarrassing. I want to hear about them. Why? Because product management is risk vs. reward. Playing safe guarantees mediocrity. Risk means occasional failure but also impact.

Here’s one from my own career: In my first assignment as a PM, I was assigned to work with an engineering org that didn’t report to my BU. I had zero formal influence. Nothing moved. My managers doubted me. In desperation, I reverse-engineered their codebase and hacked together a working feature in a few days. Suddenly, I became a hero and earned trust.

But I cut corners. The code had to run on my machine. One day, I forgot to close Chrome, and the automation took screenshots of everything including private content and published it.

What followed was 8 months of chaos, RCAs, and cleanup. Lesson learned: I did not influence the engineering team to own the code. I was a hero. I wanted the limelight.

I have also heard great answers about building clever AI/ML features based directly on user feedback - features that made perfect sense on paper. But what the PM overlooked was organizational reality: users might want that behavior, but their company policies or workflows won’t allow it.

Q5. Walk me through a product feature you built. What would you do differently?

I’m not looking for a framework. I want a story that talks about expectation vs. reality not some glossy framework:

• What sparked the idea?
• Did you see the opportunity?
• What resistance did you face?
• How did the MVP fall short?
• How did users surprise you?
• What would you change in hindsight?

The worst answers are about managing JIRA tickets. The best ones are vivid - filled with tension, misalignment, small wins, and creative bets.

I’m trying to understand if you own the product end-to-end or just facilitate others who do.

These are just 5 and I have another 5 that I will publish if there is interest.