Interview with Raj Sundaram
Sr. Director of Product

With Jefferson, you led with your execution track record — the 38%, the 21.8%, the sprints. That's the right register for a peer PM evaluating fit. Raj does not care primarily about execution metrics. He cares whether you have a point of view on where products should go and why.

Notice: same facts, completely different lens. Raj wants the so what for the world, not just for your team.

Raj likely knows you spoke with Jefferson. One of the most powerful things you can do is reference that conversation to establish continuity — it signals you're serious, you retained everything, and you're not starting from zero.

This does three things: shows preparation, gives Raj something to react to, and positions you as someone who thinks in platform relationships — exactly his frame.

At Meta, Raj led modernisation of a petabyte-scale analytics data warehouse into an AI-native, privacy-aware platform used by WhatsApp, Instagram, Reality Labs, and GenAI products. This background gives him three strong intuitions he will apply to your answers:

When in doubt, ask yourself: "Does my answer sound like it would work at 10x the scale I described it at?" If yes, you're in Raj's register. If no, reframe.

Raj has 15+ years. You have 2.5. That's not a problem you can hide — it's a dynamic you have to address directly if it comes up, because a Sr. Director will respect honesty and a specific plan far more than bravado or deflection.

"The fundamental difference is who your primary customer is. A feature product is built for end users who want a specific job done. A platform product is built for builders — developers, data scientists, product teams — who will use your platform to serve their own end users."

"Teradata's Agentic Platform is the clearest example of this: you're not building an AI agent — you're building the platform that lets AI agents reason about enterprise data reliably. The decisions you make at the platform level — what context is accessible, how trust is established, what governance hooks exist — determine what's even possible for every agent built on top."

"Three thesis statements for 3 years from now:"

What that means for Teradata: The Agentic Platform Raj is building isn't a side bet — it's the future core value proposition. ClearScape Analytics becomes the analytical engine that agents trust. VantageCloud becomes the data substrate. The question is whether Teradata makes the narrative shift fast enough from "enterprise data warehouse" to "enterprise AI operating system."

"This is ultimately a question about where your true differentiation lies and what you can defend over time."

"Developer experience (DX) is how quickly and confidently a developer can go from 'I need to solve this problem' to 'my solution is in production.' It has nothing to do with how many features the platform has — it's about friction, clarity, and trust."

"Why it matters for Teradata specifically: the data scientist who struggles with Teradata's DX will advocate for Databricks in the next architecture review. The data scientist who finds it productive will defend Teradata when the CTO asks why they're not migrating to Snowflake. DX is enterprise retention, not just a nice-to-have."

"This is the core hard problem of platform product management: your product doesn't directly generate user value — it enables others to generate user value. So your north star has to be a leading indicator of downstream success, not just platform usage."

"The key principle: your north star should make it obvious when you've succeeded and obvious when you've failed. For a platform, that usually means measuring the success rate of the things built on top of you, not just the health of the platform itself."

"A 45-year-old enterprise company is both a constraint and an asset — and the mistake is treating it as one or the other exclusively."

The constraint: Existing customer commitments slow down radical redesign. Every architectural decision carries a compatibility burden. The sales motion is built around established value propositions — shifting that narrative takes time and internal persuasion, not just good product decisions.

The asset: Teradata has something most AI startups would pay any amount to acquire: trust. Fortune 500 banks, telcos, and retailers trust Teradata with their most sensitive data. That trust is a moat that takes decades to build. Any AI platform Teradata releases starts with enterprise credibility that Databricks or a new agentic AI startup has to earn slowly."

"Pricing for agentic AI is unsolved territory — most enterprise vendors are figuring it out in real time. My framework would start with value alignment: price in a way that correlates with the value the customer extracts."

"The biggest risk isn't Snowflake or Databricks — they've been the "Teradata killer" for a decade and Teradata is still standing. The real risk is narrative displacement."

"Enterprise buyers are now asking their CDOs 'what's your AI strategy?' — not 'what's your data warehouse strategy?' If Teradata is still primarily perceived as an enterprise data warehouse company when that question is being asked, it loses the conversation before the demo. The product can be world-class, but if the category definition excludes you, the deals go to OpenAI enterprise, Google Cloud, or Databricks by default."

"The mitigation — and what Raj's team is doing — is to lead with the agentic narrative aggressively, get reference customers in AI-native use cases, and make sure the technical community (data scientists, AI engineers) is advocating for Teradata in architecture decisions."

"A traditional LLM or chatbot is reactive: you give it a prompt, it returns a response. One turn, no memory, no action in the world."

"An AI agent is autonomous and action-oriented: it receives a goal, breaks it into sub-tasks, executes those tasks using available tools (query databases, call APIs, run code, send messages), evaluates its own progress, and iterates until the goal is achieved — often without human intervention at each step."

"Teradata's Agentic Platform enables agents to do this on enterprise data — with the governance, access controls, and audit trails that make it safe to actually deploy in a bank or telco."

"An agent that reasons from general web knowledge is answering 'what does the internet think about this topic?' An enterprise-grounded agent is answering 'what is actually true in this company's data right now?' — and that distinction is the entire value proposition."

"The technical implementation: Teradata's agents likely use RAG (Retrieval-Augmented Generation) over structured SQL queries rather than over document embeddings — which is actually harder to get right, because SQL queries require the agent to understand table schemas, join relationships, and business logic. Getting this right is what Raj's team is working on."

"For Teradata's agentic use case: RAG + in-context learning for data retrieval and query generation. Fine-tuning for domain-specific behaviors like Teradata SQL dialect adherence or industry-specific output formatting."

"Most teams will be tempted to measure usage — 'agents created', 'workflows run'. That's measuring activity, not value. I'd structure metrics across three levels:"

"The 'AI replaces analysts' narrative is both misleading and actively harmful to enterprise adoption — and Teradata should position against it deliberately."

"The reality: AI agents don't replace analysts, they change what analysts do. The analyst who spent 4 hours pulling and cleaning data now reviews, validates, and acts on the agent's output in 20 minutes. That's not job replacement — that's massive capability amplification. The analyst who used to run 5 analyses per week can now run 25. Their judgment, domain knowledge, and accountability remain irreplaceable."

"The deeper truth: enterprise buyers who fear AI replacement will slow-walk adoption with governance reviews and stakeholder objections. Buyers who see AI as amplification move fast. The narrative you adopt is a product decision that affects your go-to-market velocity."

Situation: At Aasaan App, a senior leader wanted to add three new features to our B2B pitch deck based on a single large prospect's requests. The view was "if they want it, we build it."

Task: I believed we were at risk of building a custom solution for one account instead of a scalable product for our core market.

Action: Rather than just saying no, I reframed the question. I ran a quick analysis of our last 50 deals — what objections were raised, what features were requested, what actually drove conversion. The three requested features appeared in fewer than 8% of deal conversations. I presented this alongside a counter-proposal: instead of building the features, let's document them as a premium roadmap item, address this prospect's immediate needs through configuration, and run 5 more discovery calls with similar-size prospects to validate if this is a pattern before committing engineering cycles.

Result: The senior leader agreed to the discovery approach. The 5 calls revealed that 3 of the 3 features were indeed niche — we ultimately chose not to build 2 of them. One made it into the roadmap 6 months later when validated by a pattern across 10+ accounts.

"The worst response to incomplete data is paralysis — waiting for more data than you'll ever have before making a call. The second worst is pretending the uncertainty doesn't exist and making a confident call anyway."

"At Enigma, almost every decision I made was under conditions of incomplete data — we were a startup. I got comfortable being explicit about assumptions and building in review loops as a standard practice."

"I think about AI ethics not as a compliance checklist but as a product quality dimension. An AI system that makes unfair decisions, harms users, or produces unauditable outputs is simply a bad product — regardless of its accuracy metrics."

"At Teradata, where agents may be influencing decisions at Fortune 500 scale, these principles aren't abstract — they're product requirements."

"I think about this as translation work — not in the condescending sense of 'simplifying for non-technical people', but in the genuine sense of finding the shared language where everyone's actual concerns overlap."

"The test I use: can I explain this decision in a way that makes each stakeholder feel their concerns were understood, even if they didn't fully get what they asked for? If yes, the cross-functional relationship stays strong through disagreement."

"The honest answer is that 'ready to ship' is never a perfect state — it's a risk-adjusted judgment call. The question is whether the remaining imperfection poses acceptable risk to users and the business."

"A good APM executes what they're given and doesn't drop balls. A great APM changes the framing of what they're working on."

"Specifically: great APMs don't just answer 'how do I build this?' They consistently ask 'are we building the right thing?' — and they earn enough trust quickly enough that when they raise that question, people take it seriously rather than dismissing it as junior input."

"I'd rather be caught asking the right question than answering the wrong one efficiently."

"The biggest limitation I ran into was the gap between model performance on test data and adoption in the product context — what I'd call the 'last mile' problem of ML deployment."

At Aasaan, my lead-scoring model was technically strong — 46% efficiency lift in controlled testing. But it sat unused for two weeks after deployment because I'd optimised for model accuracy and completely ignored the user experience layer. Sales reps saw a score on their dashboard with no context — no explanation of what drove the score, no guidance on what to do with it."

"The lesson: ML model performance and ML product performance are different metrics. Optimising for accuracy without optimising for adoption is shipping a model that doesn't actually exist in the product."

"At Teradata, this enterprise psychology matters enormously for the Agentic Platform. The data scientist will evangelize it internally if the DX is great. But the CISO and legal team will kill the deployment if governance isn't airtight. Serving both simultaneously is the product design challenge."

"The most commercially consequential decision I made was at MyCaptain — choosing to narrow the targeting strategy rather than expand it."

When I joined, the team was selling to anyone who expressed interest — broad demographic targeting, high outreach volume, low conversion. I analysed the small set of students who had completed courses and had the best outcomes. They shared a tight profile: mid-20s, career changers from specific industries, within a decision window after a triggering event (job loss, promotion ceiling, new year)."

I recommended narrowing our entire acquisition and sales strategy to this segment and accepting lower volume in exchange for higher win rate. It was a contrarian recommendation — saying no to apparent demand is uncomfortable."

Result: Conversion rate went from 3% to 7% in 60 days. Revenue grew from ₹2L to ₹10L over 4 months. The team validated product-market fit in a segment they could actually serve well, rather than diluting across segments they served poorly."

"I'd have invested earlier in a proper product analytics infrastructure. We relied heavily on qualitative signals — user interviews, sales call notes, support tickets — because we didn't have the event tracking instrumentation to quantify user behaviour in the product itself."

"Every prioritisation decision I made at Enigma was directionally right but would have been far more defensible — and far more accurate — with product usage data behind it. I knew what users said in interviews; I didn't know what they actually did in the product minute-to-minute."

"What I'd do differently: in the first month, before working on any feature, I'd instrument the core user journeys with events — every click, every action, every drop-off. The time investment would have paid back tenfold in the quality of every subsequent decision."

"This is actually one of the things I'm most excited about at Teradata — the data infrastructure exists to make those analytics decisions precisely, not approximately."

"Pega CDH is essentially a real-time decisioning engine — it takes customer signals, applies adaptive ML models, and produces next-best-action recommendations at millisecond speed. At Enigma, I used it to build our churn prediction model that scored at-risk accounts before revenue impact appeared."

"I've been building a native Android app called FlowState — it's a personal life management system I'm developing because I couldn't find an existing app that solved the specific problem I had: building sustainable habits across multiple life dimensions simultaneously without cognitive overload."

"What makes it interesting from a PM perspective: I wrote a full PRD before writing a line of code. The PRD is grounded in neuroscience research on habit formation, and it covers six distinct modules — punctuality, wellness, social media management, smoking harm reduction, impulsive spending, and task management. Each module has its own user psychology model and success metric."

"I'm also running a personal infrastructure stack — local AI models through Ollama, n8n for automation, Portainer for container management. I do this because I believe PMs who use the technologies they're building products around have fundamentally better product instincts than PMs who only read about them."

"The honest answer is: no, I haven't worked at that scale, and I won't pretend the gap doesn't exist. I haven't designed for billions of queries per day. I haven't dealt with the governance complexity of a Meta-scale privacy-aware data platform."

"What I can offer is this: the foundational PM skills — understanding users deeply, translating between technical and business stakeholders, making defensible prioritisation decisions, measuring what matters — don't fundamentally change at scale. The vectors of difficulty change. The complexity of stakeholder management, the consequence of architectural decisions, the rigour of testing — all amplify."

"What I'd bring to that environment isn't a pretence of experience I don't have. It's a set of working instincts I've battle-tested in fast-moving, resource-constrained contexts — and genuine hunger to learn the scale-specific craft from people who've operated at Meta and HP level. That's a trade I'd make enthusiastically."

"The single principle I keep coming back to: users don't know what they want, but they always know what's frustrating them. My job is to listen hard to the frustration, resist the solution they offer, and find the underlying job-to-be-done."

"This sounds simple. In practice it's the hardest discipline in PM. The instinct to say yes to user requests — especially if they're from an important customer — is powerful. Resisting it requires confidence in the problem definition that you only get from deep user research and a clear north star. Everything else in my product process exists to serve that discipline."

"I'd look for the feature that has the highest maintenance cost relative to its actual usage. Not its stated importance — its actual usage."

"In most products there are features that were important at the time they were built, that have loud internal advocates, but that actual users have largely stopped using. These are the features that consume engineering time in every release ('we have to make sure it doesn't break') without contributing to the north star metric."

"The kill criteria: usage below X% of active users for 6+ months AND no clear roadmap reason why that changes AND the removal creates engineering headroom that funds something with higher ROI. That combination makes the business case undeniable."

"A year ago I believed that LLM accuracy was the binding constraint on AI adoption in enterprise. Build a more accurate model, get faster adoption. I've updated significantly on this."

"The binding constraint isn't accuracy — it's trust architecture. We already have models accurate enough to be useful in most enterprise contexts. What enterprises don't have is the governance framework, the audit trail, the human oversight model, and the internal accountability structure to deploy them confidently."

"This is why Teradata's approach — grounding agents in governed, auditable, structured enterprise data — is more strategically important than building a marginally better model. The problem isn't the model. The problem is everything around the model."

Product: Notion (or your actual daily tool — adapt authentically)

Change: "Notion's biggest friction for me is the absence of a reliable 'capture and organise later' mode. When I have a half-formed idea or an important piece of information, the optimal action is to capture it instantly without deciding where it goes. Currently, anything I add to Notion requires me to make a location decision at the moment of capture — which creates enough friction that I skip Notion entirely and use my phone's notes app for the initial capture."

"What I'd build: an AI-powered global inbox that takes anything — a voice note, a clipped web page, a typed thought — with zero structure decisions required. Once daily, it surfaces the captured items and suggests where they might live based on your existing Notion structure. Capture now, organise when you have the context. I believe this would change Notion from a 'deliberate workspace' tool to an 'always-on thinking partner.' The north star metric: daily active captures per user."

"You haven't asked me what excites me most about the specific problem your team is working on — the Agentic AI Platform. And I think the answer matters, because someone who's excited about data warehousing and someone who's excited about autonomous AI agents are going to show up very differently to this work."

"What excites me most: the governance problem. Everyone's building AI agents. Very few people are building them in a way that a bank's legal team would actually sign off on. The work of making agents trustworthy — explainable, auditable, scope-contained — is the hard, unresolved product problem. It's the difference between AI that's a demo and AI that's a product. That's what I want to work on."

This is the most powerful question you can ask Raj. It invites him to talk about what he actually thinks about, at the level he actually operates. It signals you're ready to engage at that level too. His answer will almost certainly be more valuable than anything in this guide, and it will make the conversation memorable for him.

You've now interviewed with both Jefferson (ClearScape Analytics PM) and Raj (Agentic AI Platform). Asking how the two products relate demonstrates you're thinking about Teradata as a system, not just two separate roles. It also shows you're already thinking like a PM who needs to understand the roadmap context, not just the immediate scope of your role.

His answer will tell you a lot about internal alignment and where the company is actually heading — information you'll need to make a good decision if you get an offer.

This question does double duty: it's a genuine rapport builder (Raj gets to reflect on his own experience) and it's self-serving information gathering for you (his answer will tell you what the real learning curve looks like if you join). It also shows intellectual curiosity about his journey, not just about the job title.

Bonus: it subtly acknowledges the experience gap without making it awkward — you're asking him to teach you something, which is exactly the posture an APM should have with a Sr. Director.

This is a different version of the "what does success look like" question — but asked in a way that's personal and direct to Raj, not generic. His answer becomes your success criteria if you get the job. It also shows you're thinking about accountability and outcomes from the first conversation, not just getting through the interview.

Ask this as your final question. It's the cleanest way to close — it leaves Raj thinking about you in the role, not just in the interview seat.