#379 The Secrets of Deploying AI in Production | Sumti Jairath, Chief Architect at SambaNova

AI agents are moving from demos into production, but many teams run into the same wall: responses that take too long and cost too much. As models grow bigger and reason through more steps before answering, every chained agent adds delay, and that adds up fast at scale. For data and AI professionals, this raises a practical question: how do you pick a model and an inference platform that keeps a multi-agent workflow fast without blowing the budget? And underneath that choice sits a bigger one — how much of the AI stack, from chips to software, should a company actually control?Sumti Jairath is Chief Architect at SambaNova Systems, where he's focused on machine learning, big data analytics, and software-defined hardware since 2017. Before that, he spent close to eight years as a Senior Hardware Architect at Oracle working on hardware acceleration for machine learning and data analytics, and earlier held processor and server design roles at Hewlett Packard and Sun Microsystems. He holds a B.Tech in Electronics and Communication Engineering from the National Institute of Technology, Kurukshetra.In the episode, Richie and Sumti explore why AI agents are slow and expensive, the architecture behind faster inference, choosing the right model and platform for a task, open-source models and sovereign AI, the full AI infrastructure stack, power efficiency, the AI data center debate, career paths in AI infrastructure, and much more.Links Mentioned in the Show:• NVIDIA's Groq acquisition, announced around GTC• Chris Olah, Anthropic• Anthropic and OpenAI's coding-agent-driven revenue growth• Ollama• Apple Mac mini• SambaNova careers• Connect with Sumti• AI-Native Course: Intro to AI for Work• Related Episode: From City Sewers to Sovereign AI with Russ Wilcox, CEO at ArtifexAINew to DataCamp? Learn on the go using the DataCamp mobile appEmpower your business with world-class data and AI skills with DataCamp for business

As agents take over more of the actual coding, the nature of technical work is shifting from producing software to judging it. Engineers increasingly spend their time specifying what should be built and then checking whether an agent's output actually solved the problem, rather than writing every line themselves. That changes what skills matter for a career in data and software: problem-solving and evaluation start to outweigh knowing today's specific tools or syntax. It also raises a harder question for teams — if agents can generate work this fast, how do you know which of it is actually worth shipping?Ledion Bitincka is Co-Founder and CTO of Cribl, where he leads the engineering organization with a first-principles approach to product delivery. Before Cribl, he was an Advanced Development Architect at Splunk, where he worked on Search-Time Schema, Hunk, and SmartStore, and before that founded Triangulus Communications. Nikhil Mungel is Head of AI R&D at Cribl, based in San Francisco, with over 15 years building distributed systems and AI teams at companies including Substack, Splunk, and ThoughtWorks — he now leads teams building LLM-powered systems for IT and security data.In the episode, Richie, Ledion, and Nikhil explore AI agent disasters and cost shocks, software telemetry fundamentals, using agents to analyze telemetry, AI-powered software factories, the shift from knowledge work to judgment work, skills for the agentic era, avoiding runaway AI spend, and measuring product value through growth metrics, and much more.Links Mentioned in the Show:• Jocko Willink's book, Leadership Strategy and Tactics: A Field Manual• Cribl• Cribl AI (Copilot)• Claude Code• AWS Graviton• NVIDIA• Connect with Ledion: LinkedIn• Connect with Nikhil: LinkedIn• AI-Native Course: Intro to AI for Work• Related Episode: AI Agents at Work: What Actually Breaks (and How to Fix It)New to DataCamp?Learn on the go using the DataCamp mobile appEmpower your business with world-class data and AI skills with DataCamp for business

Hypergrowth companies rarely scale on strategy alone — they scale on how fast decisions get made and how much risk employees feel safe taking. A recurring idea in high-growth environments is treating decisions differently depending on whether they're reversible, and rewarding people for making an impact rather than simply avoiding mistakes. For managers and individual contributors alike, that shift changes what "good work" looks like day to day. Which of your team's decisions actually need a leader's sign-off, and which ones would move faster if people just tried something and adjusted?Jon McNeill is CEO and co-founder of DVx Ventures, a venture studio that has launched 12 companies. He previously served as President at Tesla, where revenue grew from $2B to $20B in 30 months, and as COO at Lyft through its IPO. A serial entrepreneur, he's founded and sold six companies, sits on the boards of Lululemon and Asurion, and wrote The Algorithm: The Hypergrowth Formula that Transformed Tesla, Lululemon, General Motors and SpaceX.In the episode, Richie and Jon explore the algorithm behind Tesla's 10X hypergrowth, why automation should always come last, how to find and delete unnecessary process steps, building a culture of curiosity and urgency, one-way vs. two-way door decisions, small-team organizational design, changing the currency of promotion, and running effective meetings, and much more.Links Mentioned in the Show:• The Algorithm: The Hypergrowth Formula that Transformed Tesla, Lululemon, General Motors and SpaceX by Jon McNeill• Incorruptible by Eric Ries• Unreasonable Hospitality by Will Guidara• Eleven Madison Park• Jensen Huang on LinkedIn• Karim Bousta on LinkedIn• Connect with Jon• AI-Native Course: Intro to AI for Work• Related Episode: How to Thrive in a World of Continuous TransformationNew to DataCamp?Learn on the go using the DataCamp mobile appEmpower your business with world-class data and AI skills with DataCamp for business

Across the data and AI industry, infrastructure that once served dashboards and human analysts is being rebuilt to serve autonomous agents instead. That shift changes what "AI-ready data" actually means, pushing teams to rethink documentation, governance, and the semantic layer so agents pull consistent, trusted definitions rather than guessing. Day to day, this shows up as pressure to clean up gold-layer tables, eliminate duplicate metrics, and formalize business logic that used to live only in someone's head. It raises real questions: how clean does data need to be before agents can safely act on it, and who ends up owning that definition?Tristan Handy is President and Co-Founder of Fivetran + dbt Labs, the company formed by the June 2026 merger of Fivetran and dbt Labs. He founded dbt Labs in 2016 (originally as Fishtown Analytics) and spent a decade as its CEO before leading the company through the merger, and has worked in data for 23 years.In the episode, Richie and Tristan explore the dbt and Fivetran merger, building an open and modular data stack, using data to power trustworthy AI agents, the growing importance of semantic layers, how data team structures are evolving, career advice for data practitioners, context engineering for AI-driven research, and much more.Links Mentioned in the Show:• Simon Willison's blog• dbt MCP server• Apache Iceberg• Apache Polaris• LookML / Looker's semantic layer• The Vaccine Education Center (CHOP)• Connect with Tristan• AI-Native Course: Intro to AI for WorkRelated Episodes:The Data Team's Agentic Future, with Ketan Karkhanis, CEO at ThoughtSpotTowards Self-Service Data Engineering with Taylor Brown, Co-Founder and COO at FivetranNew to DataCamp?Learn on the go using the DataCamp mobile appEmpower your business with world-class data and AI skills with DataCamp for business

AI capability in mathematics jumped before most people noticed, tackling Olympiad-level problems and unsolved research questions that had resisted attack for years. But the pattern of where AI succeeds and where it stalls is uneven and worth understanding. It's much better at grinding through cases to disprove something than at constructing an elegant, original proof. Anyone working with AI in a technical field runs into this same asymmetry. Where exactly is the boundary between tasks AI can already do reliably and ones that still need human judgment and creativity?Tudor Achim is the co-founder and CEO of Harmonic, an AI company building toward mathematical superintelligence. He previously led the machine learning team at Quora and co-founded and served as CTO of the autonomous driving company Helm.ai. Under Tudor, Harmonic's Aristotle system achieved gold-medal performance at the 2025 International Math Olympiad alongside systems from OpenAI and Google DeepMind — with every proof formally verified.In the episode, Richie and Tudor explore why AI is starting to outperform humans at advanced mathematics, the shift toward formally verified proofs using the Lean language, where AI already beats humans (finding counterexamples) versus where it still falls short (building elegant proofs), how human mathematicians' roles will change, why math capability gains spill over into better AI reasoning generally, and much more.Links Mentioned in the Show:• Tudor's TED Talk: "The Path to Mathematical Superintelligence"• Aristotle, Harmonic's reasoning system• Harmonic• The Erdős Problems• American Institute of Mathematics• Rich Sutton, "The Bitter Lesson"• Connect with Tudor: LinkedIn• AI-Native Course: Intro to AI for Work• Related Episode: Why AI Agents Haven't Taken Over Knowledge Work Yet, with Jennifer Smith, CEO of Scribe (exact URL pending — episode published Aug 17, 2026, too recent to be indexed yet; confirm link on datacamp.com/podcast before publishing)New to DataCamp? Learn on the go using the DataCamp mobile appEmpower your business with world-class data and AI skills with DataCamp for business

Technology is now moving faster than the organizations trying to adopt it. A model can be tested in an afternoon, but the approval to test it can take half a year, and that gap is where most transformation budgets quietly disappear. The structures that made companies safe and predictable — layers of sign-off, centralized control, standardized processes — were built for a world where getting things wrong was expensive. That world is gone. So what actually has to change inside a company for AI to deliver value? Which habits are holding things up? And where do you start when everything needs fixing at once?Phil Le Brun is an Executive in Residence at AWS and previously spent over 25 years at McDonald's Corporation, where he was VP of Global Technology Development and International CIO. Jana Werner is an Executive in Residence at AWS, where she leads the Financial Services Practice in EMEA and advises Fortune 500 executive teams on transformation, having previously scaled a tech start-up to acquisition by HP, led digital transformation at Tesco Bank, and advised DHL on global change. Together they are the authors of The Octopus Organization: A Guide to Thriving in a World of Continuous Transformation (Harvard Business Review Press).In the episode, Richie, Phil and Jana explore why AI transformations stall, the Tin Man organization and its anti-patterns, the octopus as a model for adaptive companies, why AI adoption metrics mislead, being data informed rather than data driven, making fast reversible decisions, hiring and onboarding, embedding learning into daily work, and much more.Links Mentioned in the Show:• The Octopus Organization (book)• Through the Looking-Glass — Lewis Carroll (the Red Queen)• Goodhart's law• Annie Duke on "resulting"• Linda Hill, Harvard Business School• A Seat at the Table — Mark Schwartz• Connect with Phil• Connect with Jana• AI-Native Course: Intro to AI for Work• Related Episode: Your 90 Day Blueprint for AI Success with Charlene Li• Explore AI-Native Learning on DataCampNew to DataCamp?• Learn on the go using the DataCamp mobile app• Empower your business with world-class data and AI skills with DataCamp for business

Four years into the AI boom, headlines still promise agents that will run entire departments, yet most companies can't point to the transformation they were sold. The gap isn't intelligence — today's models are remarkably capable — it's context: no model arrives knowing how your company actually gets things done. For anyone tasked with deploying AI at work, this raises pressing questions. What does it take to turn generic intelligence into something that understands your specific operations? And why do so many well-funded AI initiatives stall before they ever reach production?Jennifer Smith is Co-Founder and CEO of Scribe, the Workflow AI platform used by more than 6 million people and 94% of the Fortune 500. Under her leadership, Scribe has surpassed $100M in ARR and raised $75M at a $1.3B valuation. Before founding Scribe, Jennifer spent three years at Greylock Partners interviewing 1,200 C-suite executives about the problems they were trying to solve, and previously worked at Coatue Management and McKinsey & Company. She holds an MBA from Harvard and a BA from Princeton.In the episode, Richie and Jennifer explore why AI agents haven't taken over knowledge work yet, harnessing institutional knowledge as "specialized intelligence," mapping enterprise workflows with LLMs, building the business case and ROI for AI transformation, balancing top-down and bottom-up change management, and the agency-driven skills that matter most in an AI-native workplace, and much more.Links Mentioned in the Show:Scribe (Jennifer's company)McKinsey & CompanyAaron Levie, CEO of Box, followed by Jennifer on XJaya Gupta, Partner at Foundation CapitalConnect with Jennifer: LinkedInAI-Native Course: Intro to AI for WorkRelated Episode: AI Agents at Work: What Actually Breaks (and How to Fix It) with Danielle Crop, EVP at WNSNew to DataCamp? Learn on the go using the DataCamp mobile app.Empower your business with world-class data and AI skills with DataCamp for business.

Software buying decisions used to be made once, by someone far removed from the people actually using the tool. That model is breaking down. Teams now expect software to work out of the box, without weeks of setup, integration, and configuration before anyone sees value. At the same time, a growing share of "users" aren't people at all — they're AI agents calling the same systems through APIs and chat interfaces. That raises a set of questions worth sitting with: what happens to user experience when the user isn't human? Can personalization and simplicity coexist, or is one always traded for the other? And as building software gets cheaper, what actually separates a good product from a cluttered one?Srini Raghavan is Chief Product Officer at Freshworks, where he leads product strategy for the company's AI-powered customer and employee experience software. He previously served as Chief Product Officer at RingCentral and SVP of Product at Five9, and holds an MBA from the University of Chicago Booth School of Business.In the episode, Richie and Srini explore the SaaS consolidation trend and why the "SaaSpocalypse" prediction missed the point, building software that works for both humans and AI agents, how MCP and modular architecture are reshaping product design, the rise of the "product builder" role replacing specialized titles, customer feedback loops and cohort-based A/B testing, judgment as the most important AI-era career skill, and much more.Links Mentioned in the Show:• Fresh Service — https://www.freshworks.com/freshservice/• Jaya Gupta's Webinar at RADAR - https://app.datacamp.com/learn/webinars/whats-next-rethinking-analytics-for-the-ai-human-era• Freddy AI Agent Studio — https://www.freshworks.com/freshservice/ai-agent-studio/• Figma Make — https://www.figma.com/make/• Cursor — https://cursor.com• NotebookLM — https://notebooklm.google/• Marc Andreessen's "Mexican standoff" comment — https://officechai.com/ai/programmer-product-manager-and-designer-roles-are-merging-into-a-single-builder-role-marc-andreessen/• Connect with Srini: https://www.linkedin.com/in/srinivasan28/• AI Tutor Course: Intro to AI for Work — https://www.datacamp.com/courses/introduction-to-ai-for-work• Related Episode: Vibe Coding and the Rise of the Non-Developer Builder with Matt Palmer, Developer Relations at ReplitNew to DataCamp? Learn on the go using the DataCamp mobile app.Empower your business with world-class data and AI skills with DataCamp for business.

Software teams are shipping faster than ever, but speed hasn't solved the oldest problem in the industry: most software still isn't very good. AI coding tools have lowered the barrier to building something, yet they haven't lowered the barrier to building something worth using. As more people who aren't trained software creators start shipping products, a new question is forming across product, design, and engineering teams: if AI can build almost anything, how do you make sure it builds the right thing, and builds it well?Todd Olson is co-founder and CEO of Pendo, the product experience platform he started in 2013. Before that, he held product and engineering roles at Rally Software, Red Hat, Cisco, and Google. He's the author of The Product-Led Organization and has led Pendo through raising over $356M in venture funding while growing to 2,300+ customers.In the episode, Richie and Todd explore why bad software still gets built, how much context AI coding agents need before they can be trusted, using behavioral data and "rage prompts" to catch what's actually frustrating users, the shift toward headless and agentic software design, how product, design, and engineering roles are splitting apart, managing one-way-door risk during AI transformation, and much more.Links Mentioned in the Show:Jeff Bezos’s one-way door / two-way door decision framework: 2015 Amazon shareholder letterHubSpot’s 2024 terms-of-service backlashRampStripeFin (Intercom’s AI agent), recently announced to be acquired by SalesforceAnthropic / Claude CodeConnect with ToddAI-Native Course: Intro to AI for WorkRelated Episode: The Data Team’s Agentic Future with Ketan Karkhanis, CEO at ThoughtSpotNew to DataCamp?Learn on the go using the DataCamp mobile app - https://www.datacamp.com/mobileEmpower your business with world-class data and AI skills with DataCamp for business - https://www.datacamp.com/business

As AI takes over more technical and routine work, the skills that set data and AI professionals apart are shifting. Raw technical ability and a high IQ still matter, but they are becoming table stakes as tools get more capable and teams get smarter. What increasingly separates people is harder to automate: communication, self-awareness, authenticity, and the ability to keep learning through failure. For anyone building a career in this space, that raises real questions. Which skills are actually worth investing in now? What holds up as AI advances? And how do you keep growing once you have already had some success?Dr. Todd Dewett is one of the world's most-watched leadership voices — an authenticity expert, bestselling author, and top LinkedIn Learning instructor whose courses have reached more than 25 million people across 100+ countries. After beginning his career at Andersen Consulting and Ernst & Young, he earned a PhD in organizational behavior at Texas A&M and spent a decade as an award-winning professor before going solo. He is a five-time TEDx speaker and the author of Show Your Ink.In the episode, Richie and Todd explore why fear quietly limits careers, treating failure as data rather than a verdict, the people skills that outlast raw IQ, learnable self-awareness, authenticity at work, using AI without losing your voice, getting better at speaking and writing, building habits, escaping the success trap, and much more.Links Mentioned in the Show:• Todd's LinkedIn newsletter (writing + his "Creswall" comic) — https://www.linkedin.com/in/drdewett/• Todd Dewett on LinkedIn Learning — https://www.linkedin.com/learning/instructors/todd-dewett• Free LinkedIn Learning access via your public library — https://www.linkedin.com/learning• Gemma Leigh Roberts, chartered psychologist — https://www.linkedin.com/in/gemmaleighroberts/• Erin Shrimpton, chartered organisational psychologist — https://ie.linkedin.com/in/erinshrimpton• Connect with Todd: https://www.linkedin.com/in/drdewett/• AI-Native Course: Intro to AI for Work• Related Episode: How to Have a Machine Learning Career in 2026 with Marina WyssNew to DataCamp?Learn on the go using the DataCamp mobile app