Tech on the Rocks

Jamie, a seasoned startup founder and former Dropbox engineer, shares insights on building distributed systems, scaling storage solutions, and the impact of AI on infrastructure and application development. Discover practical lessons from scaling Dropbox, the evolution of data storage, and how Convex is shaping the future of app development.

In this episode, we sit down with the full founding team of Wild Moose — CEO Yasmin Dunsky, CTO Roei, and VP R&D Tom Tytunovich — to explore how they’re transforming production debugging from an art into a science using AI.The trio shares their unconventional founding story — from meeting across three different cities to living together for three months in a California Airbnb to stress-test both their idea and their relationship. They discuss how they identified production debugging as a massive unsolved problem before ChatGPT even launched, recognizing that while code generation is fundamentally a text problem, debugging is a search problem that demands a completely different approach.We dive deep into Wild Moose’s “microagents” architecture — fast, highly optimized AI agents that replicate the muscle memory of senior engineers to automatically investigate production incidents in under a minute. The team explains why accuracy trumps everything in their space (wrong answers are worse than no answers when you’re debugging at 3 AM), how they navigate the speed-cost-quality triangle, and why they built a test-driven approach to validate agents against past incidents.We also get into the multi-agent vs. single-agent debate, handling multimodal observability data (logs, metrics, traces, dashboards, code), and how the rapidly evolving LLM landscape creates both opportunities and challenges for production AI systems. Plus, the team shares their favorite outage war stories — including a “WatchCat” hack and a three-month hunt for a single rogue bit.Topics covered:The Wild Moose origin story and the California Airbnb experimentWhy production debugging is a search problem, not a text generation problemMicroagents: fast, specialized AI agents for incident investigationBuilding institutional knowledge into AI — capturing engineering muscle memoryThe speed-cost-quality triangle in real-time AI systemsMulti-agent vs. single-agent architectures: when to use whatHandling multimodal observability data with LLMsThe future of AI SRE and self-healing production environmentsFavorite outage war stories from the trenchesChapters00:00 Introduction to the Wild Moose Team04:12 The Spark Behind Wild Moose08:41 Understanding the Debugging Landscape12:45 The Role of AI in Debugging17:31 Building Investigative Agents21:55 Optimizing Workflows and Feedback Loops29:12 Navigating Complexity in Software Systems33:42 Adapting to Rapid Changes in AI Technology40:02 Microagents: The Future of AI Architecture44:46 Outage Stories: Lessons from the Trenches50:49 Vision for the Future of AI in Production

In this episode, Hussain shares the story behind xorq: a “lockfile for ML pipelines” that makes notebook work easier to reproduce, debug, and ship. We talk about why the research→production path is still so manual, how schemas (and Arrow) become the contract between systems, and what it takes to run the same pipeline across engines like Snowflake and Databricks. We also dig into escape hatches for imperative code, why feature stores didn’t become the default, and how xorq fits alongside other technologies like Iceberg.Chapters00:00 Hussain's Journey in Data Science06:00 The Need for xorq: Bridging Research and Production10:38 Challenges in Machine Learning Deployment17:40 The Role of Lock Files in Data Pipelines29:51 Understanding Schema Management in Data Systems34:40 Navigating Declarative and Imperative Transformations36:39 The Developer's Journey with xorq38:34 Feature Stores vs. xorq: A Comparative Analysis43:43 The Future of Feature Stores and Machine Learning51:41 Reproducibility in Data Pipelines: xorq vs. Git-like Operations55:47 The Future of xorq and the Data Ecosystem

SummaryIn this episode of Tech on the Rocks, Kostas and Nitay sit down with Wes McKinney the creator of pandas and co-creator of Apache Arrow and Ibis, and long-time leader in the Python data ecosystem. Wes walks us through his journey from building pandas in 2008 to rethinking how we represent and move columnar data with Arrow, and why Arrow is fundamentally different from file formats like Parquet and ORC.We get into the future of data file formats, DataFusion and the new generation of query engines, the rise of open data lakes (Iceberg, Delta, Hudi), and why “big metadata” is becoming just as important as big data. Wes also shares candid thoughts on open source sustainability, how companies and infrastructure projects really survive, and how AI coding agents like Claude Code are changing the day-to-day work of software engineers, especially for complex systems work.If you care about the foundations of modern data infrastructure, or you’ve ever called import pandas as pd, this is an episode you won’t want to miss.Chapters00:00 Intro — Wes McKinney & his journey in the Python data ecosystem02:15 How pandas evolved & why UX first mattered for data science06:14 Open source sustainability, funding & the Posit model07:31 From pandas to Datapad, Cloudera & the origins of Apache Arrow and Ibis13:38 What is Apache Arrow? In‑memory columnar data, batches & schemas22:23 Inside Arrow IPC — zero‑copy, Flatbuffers & cross‑language interop24:34 Arrow vs Parquet — columnar memory format vs columnar storage format29:28 The next generation of columnar file formats & GPU‑friendly encodings36:03 Big metadata, table formats & the rise of Iceberg/Delta/Hudi43:05 Rethinking data systems: from big data to DuckDB, Rust & “no JVM” stacks54:11 DataFusion as a modular Rust query engine for modern startups57:58 Open source, the composable data stack & why infra is “AI‑resistant”01:00:07 Vibe‑coding with AI agents — using Claude Code in real projects01:09:49 AI, open source maintainers & the risks of AI‑generated contributions01:18:57 Bridging LLMs and data: ADBC, data context & the future of infra + AI

SummaryIn this conversation, the founders of Bauplan, Jacopo and Ciro, share their extensive backgrounds in AI and data infrastructure, discussing the evolution of NLP and the challenges faced in the industry. They highlight the importance of data pipelines in AI effectiveness and the complexities of building data infrastructure. The discussion also covers lessons learned from previous ventures, the shifting dynamics of the AI market, and the need for collaboration between data scientists and engineers. They emphasize the significance of simplicity in data tools and the future of data management focusing on standardization and accessibility.In this episodeBauplan was founded by experienced professionals in AI and data.Data challenges remain significant despite advancements in AI.Lessons from previous ventures inform current strategies.Building data infrastructure is complex and requires careful planning.Collaboration between data scientists and engineers is essential.Data engineering will resemble more and more software engineering.Simplicity in data tools can enhance user experience.The future of data management will focus on standardization and accessibility.If you care about making AI features shippable by regular software teams—not just data specialists—this conversation maps the terrain and the trade-offs.Chapters00:00 Introduction to Bauplan and Founders' Background02:27 The Evolution of NLP and AI Challenges05:05 Shifts in Data and AI Application07:56 Lessons from Previous Ventures10:20 The Search Market Landscape13:05 Behavioral Data's Role in Search15:52 Building Data Infrastructure vs. Applications18:22 The Complexity of Data Management21:03 Bridging the Gap Between Data Science and Engineering23:39 Challenges in Infrastructure Development29:52 Navigating the Infrastructure Landscape32:19 The Pendulum of Centralization and Decentralization34:00 The Need for Standardization in Data Infrastructure36:52 Simplifying Data Workflows40:29 Radical Simplicity in Data Management45:28 Overcoming Resistance to Change48:50 The Future of Data Abstractions and Git for Data

SummaryBrett — founder & CEO of Micro — joins Nitay and Kostas to share how he’s turning email into a knowledge graph and rebuilding CRM right inside the inbox. He traces a path from Google’s M&A and Allo product team to Clearbit and Launch House, then digs into why most “inbox zero” workflows fail, how interoperability and AI agents shift power to the interface, and what it takes to design an email experience people actually live in.What you’ll learnWhy email is a system of record—and how Micro converts threads into people, companies, attachments, tasks, and “updates”The wedge: founders’ real workflows (fundraising, hiring, sales) and why CRM belongs in the inboxProduct & UX lessons: skeuomorphic first, flexible theming (consumer vs. enterprise), and copy-the-UI-before-evolving-itM&A realities from Google: talent vs. tech vs. business acquisitions, and why culture kills most dealsBurnout and agency: why founders report less burnout than big-company rolesThe next phase: cross-app “updates” (email, LinkedIn DMs, etc.), Salesforce/HubSpot read–write, and agentic automationChapters00:00 Brett's Journey: From Consulting to Tech Innovator02:41 The Role of Strategy in Tech Companies05:16 Understanding M&A: Successes and Failures07:55 The Evolution of AI in Corporate Strategy10:26 Transitioning to Product Management13:19 Lessons from Clearbit: Culture and Growth15:50 The Impact of Burnout on Career Choices18:15 Finding Fulfillment in Entrepreneurship21:09 Navigating the B2B Landscape23:34 The Necessity of Products in a Crisis33:24 The Unexpected Layoff and New Beginnings34:39 The Launch House Experience37:16 Transforming Reality into an Accelerator39:17 The Evolution of Founders and Content Creation41:52 Introducing Micro: A New Email Experience47:02 Extracting Information for Better Workflows53:49 Integrating with Existing Ecosystems01:01:16 The Future of Email and AI

SummarySteve Klabnik has spent the last 15 years shaping how developers write code—from teaching Ruby on Rails to stewarding Rust’s explosive growth. In this wide-ranging conversation, Steve joins Kostas and Nitay to unpack the forces behind Rust’s rise and the blueprint for developer-first tooling.From Rails to Rust: How a web-framework luminary fell for a brand-new systems language and helped turn it into today’s go-to for memory-safe, zero-cost abstractions.Community as UX: The inside story of Cargo, humane compiler errors, and why welcoming IRC channels can matter more than benchmarks.Standards vs. Shipping: What Rust borrowed from the web’s rapid-release model—and why six-week cadences beat three-year committee cycles.Three tribes, one language: How dynamic-language devs, functional programmers, and C/C++ veterans each found a home in Rust—and what they contributed in return.Looking ahead: Steve’s watch-list of next-gen languages (Hylo, Zig, Odin) and the lessons Rust’s journey holds for anyone building tools, communities, or startups today.Whether you’re chasing segfault-free code, dreaming up a new PL, or just curious how open-source movements gain momentum, this episode is packed with insight and practical takeaways.Chapters00:00 Introduction and Personal Connection00:59 Journey from Ruby on Rails to Rust02:21 Early Programming Experiences and Interests07:20 Community Dynamics in Programming Languages13:59 The Importance of Community in Open Source14:37 How Ruby on Rails and Rust Built Their Communities21:44 Standardization vs. Unified Development Models30:55 Community Debt in Programming Languages36:24 Release Cadence vs. Feature Development37:36 Rust's Unique Selling Proposition43:30 Attracting Diverse Programming Communities52:31 The Future of Systems Programming Languages

SummaryJosh Howard, Senior Engineering Manager at Cloudflare, joins Kostas and Nitay to discuss Cloudflare's innovative serverless platform, Durable Objects, and Workers. Learn how Cloudflare enables developers to build stateful applications with global scale, consistency, and simplicity at the network edge.Chapters00:00 Introduction and Background02:01 Journey into Storage Systems04:24 Cloudflare's Evolution and Developer Platform06:29 Understanding Durable Objects08:57 Durable Objects in Modern App Development11:18 Use Cases for Cloudflare's Developer Platform13:36 Building Agents and Real-Time Applications16:19 Developer Experience and Migration Strategies25:09 Exploring Workflow Systems: OLAP vs Applications26:47 Cloudflare's Development Platform: Future Offerings for Data Professionals28:42 Transitioning from Data Processing to Application Development31:37 The Impact of LLMs on System Design33:44 Serverless Platforms: Challenges and Limitations40:01 Future Directions: Cloudflare's Storage Relay Service and Global ExpansionClick here to view the episode transcript.

SummaryIn this episode, Erik reflects on his long and storied tech career—from the days of punch cards to founding multiple startups, including a stint at Splunk. At 61, he offers a unique perspective on how the industry has evolved and shares candid insights into what it takes to build a successful company. He discusses the evolution from building simple tools to creating comprehensive solutions and eventually platforms, emphasizing the importance of starting with a “hammer”—a focused, simple tool—before scaling to a broader offering. Eril introduces his concept of the “physics of business,” a framework for understanding go-to-market dynamics, pricing, and the critical role of brand in differentiating a product in a crowded market. He also touches on the challenges of product-led growth, the importance of achieving a strong “K value” (viral or network effects), and the pitfalls of allowing short-term quarterly pressures to derail long-term vision. Toward the end, he hints at his current project, Bestimer, which aims to apply lessons from his past ventures and leverage modern AI to tackle a massive, data-intensive problem.Chapters00:00 Erik's Journey Through Tech History04:06 The Philosophy of Designing for Success09:49 Understanding the Physics of Business14:29 Timing and Luck in Startups18:09 Lessons Learned from Splunk23:30 The Power of Brand in Business28:02 Leveraging AI for Brand Development32:04 The Resilience of Splunk36:45 Building a Competitive Edge37:28 From Tool to Solution40:59 The Importance of Onboarding44:32 Navigating Growth and Market Fit51:11 Innovating with AI: The Next Chapter

SummaryIn this episode, Gilad Kleinman, co-founder of Epsio, shares his unique journey from PHP development to low-level kernel programming and how that evolution led him to build an innovative incremental views engine. Gilad explains that Epsio tackles a common challenge in databases: making heavy, complex queries faster and more efficient through incremental materialization. He describes how traditional materialized views fall short—often requiring full refreshes—and how Epsio seamlessly integrates with existing databases by consuming replication streams (CDC) and writing back to result tables without disrupting the core transactional system. The conversation dives into the technical trade-offs and optimizations involved, such as handling stateful versus stateless operators (like group-by and window functions), using Rust for performance, and the challenges of ensuring consistency. Gilad also contrasts Epsio’s approach with streaming systems like Flink, emphasizing that by maintaining tight integration with the native database, Epsio can offer immediate, up-to-date query results while minimizing disruption. Finally, he outlines his vision for the future of incremental stream processing and materialized views as a means to reduce compute costs and enhance overall system performance.Chapters00:00 From PHP to Kernel Development: A Journey07:30 Introducing Epsio: The Incremental Views Engine10:56 The Importance of Materialized Views15:07 Understanding Incremental Materialization19:21 Optimizing Query Performance with Epsio24:53 Integrating Epsio with Existing Databases27:02 The Shift from Theory to Practice in Data Processing29:42 Seamless Integration with Existing Databases32:02 Understanding Epsio Incremental Processing Mechanism34:46 Challenges and Limitations of Incremental Views36:49 The Complexity of Implementing Operators39:56 Trade-offs in Incremental Computation41:21 User Interaction with Epsio43:01 Comparing EPSIO with Streaming Systems45:09 Architectural Guarantees of Epsio50:33 The Future of Incremental Data Processing

SummaryIn this episode, Viktor Kessler shares his journey and insights from his extensive experience in data management—from building risk management systems and data warehouses to working as a solutions architect at MongoDB and Dremio, and now co-founding a startup.Initially exploring data mesh concepts, Viktor explains how real-world challenges—such as the disconnect between technical data models and business needs, inconsistent definitions across departments, and the difficulty in managing actionable metadata—led him and his co-founder to pivot toward building a lake house solution. His startup is developing Lakekeeper, an open source REST catalog for Apache Iceberg, which aims to bridge the gap between decentralized data production and centralized metadata management. The conversation also delves into the evolution of data catalogs, the necessity for self-service analytics, and how creating consumption-ready data products can transform data functions from cost centers into profit centers. Finally, Viktor outlines ways for interested listeners to get involved with the Lakekeeper community through GitHub, upcoming meetups, and a dedicated Discord channel.Chapters00:00 Introduction to Viktor Kessler and His Journey04:57 Transitioning from Data Mesh to Lake House09:15 Understanding Data Mesh: Pain Points and Solutions13:47 The Role of Metadata in Data Management18:16 The Evolution of Catalogs and Metadata Management28:14 Stabilizing the Consumption Pipeline31:18 Centralizing Metadata for Decentralized Organizations37:09 Bridging the Gap: Technical and Business Perspectives43:17 Rethinking Data Products and Consumption50:45 Finding Balance: Control and Flexibility in Data Management

SummaryIn this episode, Apurva Mehta, co-founder and CEO of Responsive, recounts his extensive journey in stream processing—from his early work at LinkedIn and Confluent to his current venture at Responsive. He explains how stream processing evolved from simple event ingestion and graph indexing to powering complex, stateful applications such as search indexing, inventory management, and trade settlement. Apurva clarifies the often-misunderstood concept of “real time,” arguing that low latency (often in the one- to two-second range) is more accurate for many applications than the instantaneous response many assume. He delves into the challenges of state management, discussing the limitations of embedded state stores like RocksDB and traditional databases (e.g., Postgres) when faced with high update rates and complex transactional requirements. The conversation also covers the trade-offs between SQL-based streaming interfaces and more flexible APIs, and how Responsive is innovating by decoupling state from compute—leveraging remote state solutions built on object stores (like S3) with specialized systems such as SlateDB—to improve elasticity, cost efficiency, and operational simplicity in mission-critical applications.Chapters00:00 Introduction to Apurva Mehta and Streaming Background08:50 Defining Real-Time in Streaming Contexts14:18 Challenges of Stateful Stream Processing19:50 Comparing Streaming Processing with Traditional Databases26:38 Product Perspectives on Streaming vs Analytical Systems31:10 Operational Rigor and Business Opportunities38:31 Developers' Needs: Beyond SQL45:53 Simplifying Infrastructure: The Cost of Complexity51:03 The Future of Streaming ApplicationsClick here to view the episode transcript.

In this episode, we chat with David Jayatillake, VP of AI at Cube, about semantic layers and their crucial role in making AI work reliably with data. We explore how semantic layers act as a bridge between raw data and business meaning, and why they're more practical than pure knowledge graphs. David shares insights from his experience at Delphi Labs, where they achieved 100% accuracy in natural language data queries by combining semantic layers with AI, compared to just 16% accuracy with direct text-to-SQL approaches. We discuss the challenges of building and maintaining semantic layers, the importance of proper naming and documentation, and how AI can help automate their creation. Finally, we explore the future of semantic layers in the context of AI agents and enterprise data systems, and learn about Cube's upcoming AI-powered features for 2025.00:00 Introduction to AI and Semantic Layers05:09 The Evolution of Semantic Layers Before and After AI09:48 Challenges in Implementing Semantic Layers14:11 The Role of Semantic Layers in Data Access18:59 The Future of Semantic Layers with AI23:25 Comparing Text to SQL and Semantic Layer Approaches27:40 Limitations and Constraints of Semantic Layers30:08 Understanding LLMs and Semantic Errors35:03 The Importance of Naming in Semantic Layers37:07 Debugging Semantic Issues in LLMs38:07 The Future of LLMs as Agents41:53 Discovering Services for LLM Agents50:34 What's Next for Cube and AI Integration

In this episode, we chat with Yaron Hadad, a fascinating individual who transitioned from theoretical physics to entrepreneurship. We explore his groundbreaking work on black holes and gravitational waves, and learn about the Ramanujan Machine - an algorithmic system he helped develop that discovers new mathematical formulas and democratizes mathematical research. We'll hear about the scientific community's mixed reactions to this innovative approach. The conversation then shifts to his work with Neutrino, a company he founded that uses AI and continuous monitoring devices to understand how food affects individual health. We delve into the complexities of nutrition science, the challenges of processing multiple data streams, and the future of personalized health monitoring. Throughout the episode, Yaron shares insights on bridging theoretical research with practical applications, and the role of AI in advancing both pure mathematics and healthcare.00:00 Yaron Hadad's Journey: From Physics to AI in Healthcare04:50 The Complexity of Einstein's Equations and Their Solutions10:12 AI in Mathematics: The Ramanujan Machine and Conjectures15:41 Navigating Criticism: The Scientific Community's Response to Innovation29:24 The Impact of Algorithms in Mathematics35:30 The Planck Machine: A New Approach41:15 Neutrino: A Personal Journey in Nutrition50:11 Connecting Food Complexity to Health Metrics

In this episode, we chat with Philippe Noël, founder of ParadeDB, about building an Elasticsearch alternative natively on PostgreSQL. We explore the challenges and benefits of extending PostgreSQL versus building a separate system, diving into topics like full-text search, faceted analytics, and why organizations need these capabilities. We discuss the emerging bring-your-own-cloud deployment model, the state of the PostgreSQL extension ecosystem, and what makes a truly production-ready database extension. Philippe shares insights on the future of search technology and how recent AI developments are actually increasing the demand for traditional search capabilities. The conversation also covers the misconceptions around PostgreSQL's scalability and the trade-offs between multi-tenant and single-tenant architectures in modern data infrastructure.Chapters00:00 Introduction to ParadeDB and Its Mission06:35 User-Facing Search and Analytics11:45 The Role of Postgres in Modern Data Solutions17:30 Future of Multimodal Databases31:04 The Rise of Fintech and Data Integrity36:36 Deployment Models: BYOC and Control Plane43:41 The Evolution of Cloud Infrastructure and Serverless Databases49:38 The Future of Search and Community EngagementClick here to view the episode transcript.

In this episode, we chat with Ben, founder of Espresso AI, about his journey from building Excel Python integrations to optimizing data warehouse compute costs. We explore his experience at companies like Uber and Google, where he worked on everything from distributed systems to ML and storage infrastructure. We learn about the evolution of his latest venture, which started as a C++ compiler optimization project and transformed into a system for optimizing Snowflake workloads using ML. Ben shares insights about applying LLMs to SQL optimization, the challenges of verified code transformation, and the importance of formal verification in ML systems. Finally, we discuss his practical approach to choosing ML models and the critical lesson he learned about talking to users before building products.Chapters00:00 Ben's Journey: From Startups to Big Tech13:00 The Importance of Timing in Entrepreneurship19:22 Consulting Insights: Learning from Clients23:32 Transitioning to Big Tech: Experiences at Uber and Google30:58 The Future of AI: End-to-End Systems and Data Utilization35:53 Transitioning Between Domains: From ML to Distributed Systems44:24 Espresso's Mission: Optimizing SQL with ML51:26 The Future of Code Optimization and AIClick here to view the episode transcript.

In this episode, we chat with David Mytton, founder and CEO of Arcjet and creator of console.dev. We explore his journey from building a cloud monitoring startup to founding a security-as-code company. David shares fascinating insights about bot detection, the challenges of securing modern applications, and why traditional security approaches often fail to meet developers' needs. We discuss the innovative use of WebAssembly for high-performance security checks, the importance of developer experience in security tools, and the delicate balance between security and latency. The conversation also covers his work on environmental technology and cloud computing sustainability, as well as his experience reviewing developer tools for console.dev, where he emphasizes the critical role of documentation in distinguishing great developer tools from mediocre ones.Chapters00:00 Introduction to David Mytton and Arcjet07:09 The Evolution of Observability12:37 The Future of Observability Tools18:19 Innovations in Data Storage for Observability23:57 Challenges in AI Implementation31:33 The Dichotomy of AI and Human Involvement36:17 Detecting Bots: Techniques and Challenges42:46 AI's Role in Enhancing Security47:52 Latency and Decision-Making in Security52:40 Managing Software Lifecycle and Observability58:58 The Role of Documentation in Developer ToolsClick here to view the episode transcript.