Scaling AI in MarTech: Agentic Workflows, High-QPS Systems, and Human-in-the-Loop Leadership with Mohit Nagpal

About the Show Mind Meets Machine is a podcast produced under the Podhub Network and hosted by Avik Chakraborty. The show explores the intersection of human intention, business growth, and technology. It provides clear strategy for creators, founders, and service providers who want to build sustainable digital systems without losing human connection. About the Guest Mohit Nagpal is a senior software engineering leader at Uber with 15 years of experience in distributed systems, high-throughput backend architecture, and Marketing Technology (MarTech). He specializes in scaling marketing automation, optimizing real-time experiment pipelines processing tens of thousands of queries per second (QPS), and integrating agentic AI workflows into enterprise platforms. To learn more about Mohit or connect with him directly, visit his official LinkedIn profile: https://www.linkedin.com/in/mohitnagpal/   Why You Must Listen Integrating Generative AI into enterprise systems requires more than plugging in an API. This conversation offers a practical look at scaling AI infrastructure across millions of users without sacrificing system reliability or brand safety. Mohit provides concrete architectural insights into handling 50,000+ QPS workloads, balancing offline caching with real-time personalization, and structuring human oversight to prevent production failures. Key Takeaways Forced AI adoption in non-essential areas wastes engineering resources; teams must evaluate cost vs. ROI before replacing working systems. Human-in-the-loop (HITL) verification is essential in automated marketing to protect brand identity, maintain legal compliance, and prevent AI hallucinations. Agentic workflows accelerate campaign preparation—including audience segmentation, budget documentation, and content translation—reducing rollout timelines from months to weeks. Running real-time user personalization at high throughput (50,000+ QPS) requires balancing parameter complexity against execution latency through intelligent caching. Early-career engineers should focus on building real-world projects with modern AI developer tools to demonstrate practical problem-solving capability.   💬 Want to come on the show? Be a Guest: https://www.podhub.club/beaguest 🌐 Explore the Network & Community: Full Network: https://www.podhub.club/podcastnetwork Newsletter: https://healthymindbyavik.substack.com/ LinkedIn Community: https://www.linkedin.com/groups/14498362/ This isn't self-help. It's self-honesty. 💼 Business & Sponsorships: Sponsor Our Show: https://www.podhub.club/sponsors Check Our Services: https://www.podhub.club/services Connect & Follow Avik & Healthy Mind: 📝 Medium Articles: https://medium.com/@contentbyavik 📝 MYND Studio Publication: https://medium.com/mynd-studio 📹 YouTube Channel: https://www.youtube.com/@healthymindbyavik?sub_confirmation=1 📸 Instagram: https://www.instagram.com/healthyminds.pod/ 📘 Facebook: https://www.facebook.com/podcast.healthymind 💼 LinkedIn Company Page: https://www.linkedin.com/company/healthymindbyavik 💼 Avik's LinkedIn Profile: https://www.linkedin.com/in/healthypodminds 🐤 X / Twitter: https://twitter.com/podhealthclub 📌 Pinterest: https://www.pinterest.com/Avikpodhealth/ 💬 Reddit: https://www.reddit.com/user/Podcastbyavik/ 🧵 Threads: https://www.threads.com/@healthyminds.pod 📌 Disclaimer: This episode is for educational and informational purposes only. Guest views are personal and do not represent the host or Healthy Mind by Avik™. The Network does not verify or endorse guest statements. Nothing here is medical, legal, financial, or professional advice; please consult a qualified professional. Engage critically. Third-party content referenced under fair use. Guests are responsible for their own statements. Contact Us: https://www.podhub.club/contact Full Disclaimer: https://www.podhub.club/terms-privacy/standard-podcast-disclaimer By listening, you accept this disclaimer in full.   Transcript 1 00:00:00,000 --> 00:00:15,000 Host (Avik): Dear listeners, there is a quiet moment before a team ships something new and uncertain—a technology that did not exist in production a year ago, or a workflow that could either collapse under scale or quietly change how millions of decisions are made. 2 00:00:15,000 --> 00:00:53,000 Host (Avik): Today, we are with someone who has been in that moment more than once and chose to lead his team through it anyway. Welcome to Mind Meets Machine. I am your host, Avik, and today we have a very lovely guest, Mohit Nagpal. Welcome to the show. 3 00:00:53,000 --> 00:01:31,000 Guest (Mohit): Thanks, Avik. Great to be on your show. I have been working in the industry for 15 years, recently at Uber as a software engineer in Marketing Technology, leading the shift from traditional workflows to AI-based workflows scaling to millions of users. 4 00:01:31,000 --> 00:02:00,000 Host (Avik): There is a lot of noise around generative AI right now—both hype and skepticism. What is the biggest misconception engineering leaders hold when adopting it? 5 00:02:00,000 --> 00:03:30,000 Guest (Mohit): The biggest misconception is failing to evaluate where GenAI fits well versus where it does not. The non-trivial cost of GenAI means force-fitting it everywhere makes little sense. Organizations are realizing the need to be more selective about where AI integration truly adds value. 6 00:03:30,000 --> 00:03:50,000 Host (Avik): When introducing AI agents into marketing workflows, what initial reservations did your team raise, and how did you address them? 7 00:03:50,000 --> 00:05:14,000 Guest (Mohit): Marketing requires brand alignment, storytelling, legal compliance, and regional tone. The primary concern was whether AI would automatically output content that violates brand guidelines. We address this by maintaining a human-in-the-loop setup to review and verify generated output before launch. 8 00:05:14,000 --> 00:05:50,000 Host (Avik): Adoption is fundamentally a trust challenge rather than purely a technical one. You mentioned reducing setup timelines by 4x using AI agents. What was broken in the old workflow that made this possible? 9 00:05:50,000 --> 00:09:27,000 Guest (Mohit): Traditionally, launching a major global campaign—such as targeting 250 million users for a holiday push—took over a month. Steps included budget approvals, data analysis, audience selection, creative asset design, and multi-language translation. AI agents compress tasks like audience queries, budget documentation, and localization from days into hours, reducing launch timelines from months to weeks. 10 00:09:27,000 --> 00:09:45,000 Host (Avik): Running experimentation across thousands of queries per second reveals system bottlenecks. What challenges emerge at that scale? 11 00:09:45,000 --> 00:12:19,000 Guest (Mohit): When serving 50,000 to 100,000 QPS across billions of notifications, evaluating multiple personalization parameters per user in real-time creates systemic latency challenges. Engineering teams must strike a balance between deep personalization and system throughput through smart offline processing and caching. 12 00:12:19,000 --> 00:12:56,000 Host (Avik): What changes for a mid-level engineer who has never interacted with generative AI models directly? 13 00:12:56,000 --> 00:15:32,000 Guest (Mohit): Previously, building ML pipelines required close pairing with specialized data science teams. Today, GenAI developer tooling lowers that entry barrier, allowing software engineers to conduct research, prototype features, and handle initial model interactions independently. 14 00:15:32,000 --> 00:17:22,000 Guest (Mohit): Because LLM outputs carry no inherent accountability and can behave unpredictably in production, maintaining a human-in-the-loop framework is vital for debugging edge cases and maintaining system reliability. 15 00:17:22,000 --> 00:17:39,000 Host (Avik): What can early-career engineers start doing now to prepare for leading technological transitions in their teams? 16 00:17:39,000 --> 00:19:33,000 Guest (Mohit): Focus on building real-world projects. Pick a practical domain, leverage modern AI tools, and solve tangible problems. Demonstrating practical problem-solving experience sets you apart during engineering discussions and team evaluations. 17 00:19:33,000 --> 00:19:39,000 Host (Avik): Where can listeners connect with you online? 18 00:19:39,000 --> 00:19:54,000 Guest (Mohit): Listeners can connect with me directly on LinkedIn. 19 00:19:54,000 --> 00:20:38,000 Host (Avik): All links will be included in the show notes. Technology is rarely the hardest part of change—trust, patience, and clear leadership carry equal weight. Thank you, Mohit, and thank you to our listeners.