Episodes

  • Ep. 23 - How AI is Rewriting Analytics
    Nov 18 2026

    AI is changing more than the tools people use to do their jobs. In Yuri Gubin’s experience, it is affecting both how his team develops solutions and what those solutions can make possible for clients.

    Gubin describes his team’s work as developing solutions that help clients innovate. AI is now part of that work: it supports code creation and architectural work, and it is influencing the kinds of solutions the team creates.

    The change, then, is not limited to using AI to complete familiar tasks differently. It also extends to how teams approach the work of building solutions.That same shift is visible in data.

    Gubin points out that working with data is not only about making information available for queries or creating dashboards. Those tools can help present and access information, but he sees a broader opportunity: enabling more people to work directly with data, explore ideas through faster experiments and make information more accessible to stakeholders who need it for decision-making.

    This broader access matters because experimentation can help teams learn and assess possibilities sooner. When more people can engage with data and test ideas, data work can support decision-making in new ways not only by presenting results, but by helping people explore what they might mean.

    For Gubin, AI’s potential lies in this combination: using it in the development process while also making data more accessible and useful across an organization. The aim is not simply to produce code or dashboards more quickly. It is to help teams and stakeholders do more with the information they have, experiment faster and create new possibilities for their work.



    Get full access to My Data Guest at mydataguest.substack.com/subscribe
    Show More Show Less
    34 mins
  • Ep. 24 - Navigating the Open Source Data Landscape
    Nov 11 2026
    Some guests need a long introduction. Marco Gorelli is one of them, so I’ll keep it short: mathematician, software engineer at Quansight Labs, former pandas maintainer, Polars contributor, and the creator of Narwhals, a library that gives you one API on top of the major dataframe libraries.We first met at PyCon Italy, when Narwhals was only a few months old, and we’ve run into each other at conferences since. This time we sat down properly and talked about the stuff that usually only comes up in a conference hallway: why some APIs win, how you get anyone to notice a new open source project, and what maintainers are dealing with now that coding agents can open pull requests.Here are the highlights.From pure maths to pandas maintainerMarco studied mathematics, followed by a master’s in maths and the foundations of computer science. Despite the name, it involved no programming at all, “much to some recruiters’ disappointment” when they tried to place him in very applied jobs.He went into data science around 2017–2018, when the field was booming. He pointed out that the way people talked about data science back then sounds a lot like the way people talk about AI today, which in turn sounds like how they talked about big data a few years earlier. (When was the last time you heard someone say “big data”?)The way into open source was very ordinary. He hit a bug in pandas at work and thought it might be fun to fix it. Then COVID hit, he had extra time, and he kept going. Eventually he became a pandas maintainer, and that led to a job at Quansight Labs working on open source full time.Why Polars won by not copying pandasAt Quansight, Marco got into Rust and Polars, a dataframe library written in Rust that avoids many of pandas’ historical design decisions.It’s easy to forget how recently people “would look at me as if I had three heads” if he mentioned Polars alongside pandas. Early on, the critics said nobody would use a dataframe library without the pandas index or the pandas API. Polars was adopted widely anyway. Meanwhile, several projects that copied pandas exactly, only faster, have mostly disappeared.Marco’s explanation: Ritchie Vink, the creator of Polars, treated not following the pandas API as a feature and not a limitation.> “Yes, there is a learning curve, but once you’re over that learning curve, everything is very consistent. Once you get used to the Polars way of thinking, it is a pleasure to work with.”His practical advice: the best time to try Polars is when you’re starting a new project. Don’t try to recreate your pandas habits. Learn the Polars way of thinking from the start and you get the benefits.Why pandas struggles to keep upI asked whether many of the complaints about pandas come from the fact that it has been around so long, through data mining, big data, machine learning, deep learning and now generative AI.Marco agreed and pointed to hardware. pandas was designed around 2008, when 16-core CPUs weren’t something everyone had. Today they’re common and cheap to rent. The problem is that the pandas API doesn’t make it obvious what can run in parallel. People say “pandas should just do things in parallel”, but the API doesn’t make clear how that would work.Polars makes it explicit. Each expression inside with_columns runs independently, so they can all run multi-threaded.> “The world has certainly changed around pandas, but it’s not easy for pandas to change itself at the same pace.”What you gain by going from user to contributorMarco named two things.1. You appreciate the invisible work. pandas is run entirely by volunteers, with the occasional grant. Polars has a company behind it, but much of what goes into the free tool isn’t driven by money. It’s driven by people who want to make data work better.2. You learn API design the hard way. Users often think there should be a built-in function for their use case. But if a library added a function for every use case, the API would become impossible to navigate, and every function would have 20 parameters. As the old pandas maintainer joke goes:> “There is no problem that cannot be solved by adding an extra keyword argument to read_csv.”(It’s probably past 50 by now.) A good tool has to be fairly minimal. It can’t solve everyone’s problem, but it can give you the building blocks to solve your own.Narwhals: copying Polars, not pandasI’ve always found it fascinating that something as simple as mean behaves differently across libraries. Marco gave an example right away: missing values. Most tools skip nulls by default, but the way you configure that differs between PyArrow, pandas, Polars and SQL.That’s the kind of problem Narwhals solves. Marco started it about two and a half years ago as a unified dataframe API, aimed mainly at library maintainers who want to accept different kinds of dataframe inputs.He had three options: copy pandas, copy another library, or invent a ...
    Show More Show Less
    41 mins
  • One year... And what's coming next!
    Oct 21 2026

    A year ago, Rosaria and I started My Data Guest with a simple goal: understand what was happening in AI by talking to the people building it. Most of that first season was about the how. How a transformer works, how a large language model actually processes context, how this new wave of tools was different from the data science we were used to.

    After a two-month break, we’re back. And the world we’re coming back to looks very different.

    My Data Guest is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

    The novelty wore off, and that’s a good thing

    Something changed in the last twelve months. New models still come out every few weeks, but the architecture underneath is mostly the same: bigger, faster, with a few more tricks. The excitement of discovery has quietly turned into something more practical. Rosaria called it a plumbing exercise, and I think she’s right.

    I noticed it in my own work. A year ago I was building an n8n workflow to post My Data Guest content on LinkedIn, and debugging it meant digging through logs. Today I don’t even open the logs. Claude Code is connected to n8n through an MCP server and looks into the problem for me. What surprised me wasn’t the capability itself, but how fast it started to feel normal.

    That’s the shift we want to follow this season.

    What we’re focusing on next

    When almost anything can be connected to anything, the question stops being can we build it and becomes what should we build, and how do we keep it under control once it’s running. We went from fear of missing out to being overwhelmed by options.

    Two themes keep coming up in conversations, posts, and articles:

    Production. How do you actually deploy an agentic system? I’ve come to see agents as distributed systems with an LLM in the middle, and that means a lot of what we already know from system design and cybersecurity still applies. The hard part is translating it.

    Legacy integration. Nobody gets to throw their existing stack away. Companies have years of systems in production, and now those systems have to talk to something completely new. It reminds me of the COBOL days. The old world doesn’t disappear; it has to coexist.

    And underneath both of these sits the topic I expect to dominate the year: security and control. LLMs aren’t just calling tools anymore, they’re writing their own and executing them. That only works safely inside a proper sandbox. We’ll talk about why, and we’ll dig into a recent story that shows what happens when a sandbox isn’t enough, trying to separate what actually happened from the marketing around it.

    What’s coming on the channel

    * Interviews stay free and open, as always. The first one of the season is with Marco Gorelli, and the story of how we met involves a Python package and a few pull requests.

    * New Learning Bytes, same format as before but more hands-on, built around deployment, safety, and what can go wrong before an agent hits production.

    * Courses and masterclasses for paying members: structured, multi-hour learning that takes you from zero to a working product, possibly at different levels, including for people coming from outside tech.

    * A Telegram group for AI news and channel updates. This one is happening for sure.

    There are a few more ideas in the episode, including a possible collaboration and a plan we’ve now announced twice (maybe this is the year).

    My Data Guest is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

    One more thing

    Last year we made a prediction: as more content becomes AI-generated, people would start valuing the human touch more. Real effort, real imperfection. Rosaria checked whether it came true, and the conversation took a turn I didn’t expect, from theatres to Hamburg’s new opera house to The Polar Express.

    You’ll have to watch for that one.

    If you want to be a guest, or you have a course you’d like us to build, reach out on LinkedIn or Substack. And thanks for staying with us for another year.



    Get full access to My Data Guest at mydataguest.substack.com/subscribe
    Show More Show Less
    33 mins
  • Ep. 21 - AI Agents: The Gap Between Demo and Production
    Oct 14 2026

    My Data Guest is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

    After a two-month break, My Data Guest is back with Bilge Yücel, Senior DevRel Engineer at deepset. This one’s less about a product and more about the shifts happening underneath the AI hype: how people learn to code now, what documentation means when agents read it instead of humans, and how you measure trust in a community that can’t be reduced to a metric.

    Bilge’s own path wasn’t linear. She started as a frontend engineer in Turkey, got curious about how tools like Google Translate actually worked, and left a stable job to study NLP, pre-ChatGPT. She didn’t want to be a researcher chasing a PhD. She wanted to build things people could use, which is what eventually pulled her into developer relations.

    That personal arc sets up the bigger question in the episode: what happens to problem-solving skills when AI writes the code? Bilge points out that her generation learned by writing C by hand, wrestling with memory, debugging without a model to lean on. Today’s students are allowed to use generative AI for coursework. The skill isn’t gone, it’s shifted, from writing code line by line to knowing how to direct, evaluate, and trust an agent’s output.

    That shift shows up again later in a different form. Stack Overflow traffic has cratered since ChatGPT. People don’t search anymore, they ask an agent. So documentation itself has had to adapt: not just written for humans, but structured so coding assistants retrieve the current version instead of hallucinating outdated syntax. It’s a small but telling example of how much invisible infrastructure work has to happen just to keep information trustworthy in an agent-first world.

    The conversation also gets into the murkier side of community-facing work. How do you measure whether a DevRel role is succeeding? Pip installs and GitHub stars are noisy signals at best, mostly useful for spotting seasonality, not proving impact. Bilge’s answer is blunter: trust is the real currency, and trust doesn’t show up on a dashboard. You collect it in anecdotes, in feedback forms, in the fact that someone comes up to you at a conference and says a tool genuinely helped them. It’s unsatisfying from a KPI standpoint and probably the most honest answer available.

    There’s good practical advice here too, for anyone who’s ever wanted to speak at a conference. Developers, she says, should look at whatever they just shipped, even something small, and find the story in it. DevRels operate differently: submitting talks a year out based on a roadmap, sometimes with no slides and no code yet written, just a belief they’ll build it in time.

    And a broader, honest tension runs through the whole episode: what does neutrality look like when you work for a company that needs to make money? Bilge’s answer is that her job isn’t to sell, it’s to be trusted, which sometimes means telling someone a tool isn’t the right fit for them.

    Worth a full listen if you care about how technical communities actually function, and what’s really being lost or gained as AI reshapes how we learn and document things.

    My Data Guest is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.



    Get full access to My Data Guest at mydataguest.substack.com/subscribe
    Show More Show Less
    43 mins
  • Ep. 20 - Beyond AI Janitors
    Jul 29 2026

    Discover how top universities are transforming teaching methods to prepare students for a future where AI is an essential co-pilot, not a crutch. In this eye-opening episode, seasoned professors from UVA’s business and commerce schools reveal the revolutionary shift from traditional teaching to a dynamic, AI-augmented learning experience.

    You’ll hear from Ryan Wright, a UVA professor working to integrate AI into learning and research, Tim Laseter, who reflects on how the case method is evolving, and Keith McCormick, who brings a practical machine-learning perspective on AI’s real impact in organizations. Together, they unpack how students should be using AI as a co-creator, why judgment and critical thinking matter more than ever, and how educators can preserve genuine learning in an AI-driven world.

    You'll learn how case method classrooms are becoming interactive labs where students collaborate with AI tools, analyze real-world data, and develop judgment skills that no AI can replace.

    Ryan Wright wonders how we can teach the students not just to be ‘AI janitor,’ cleaning up and refining outputs, but rather co-creator partners of AI. Tim Laseter explores how to harness AI in experiential learning and live cases, making students thrive in an uncertain future. Keith McCormick highlights the importance of understanding organization memory and how AI can augment, rather than replace, the uniquely human skills of empathy and judgment.

    In this episode, we also break down concrete tactics for educators and students alike: from using AI-driven peer reviews and meta-analysis to cultivating critical thinking through iterative questioning. The conversation then tackles the delicate balance of assessment in a world where AI is ubiquitous: should exams be oral, project-based, or involve real-time interrogation of AI outputs?

    Beyond individual skills, the episode underscores a profound insight: how should we shape creativity in students? Possibly, a liberal arts foundation is the best strategy for future-proofing careers in the AI era. As AI accelerates change, the ability to see beyond AI-generated answers, ask the right questions, and think critically becomes invaluable.

    This episode is essential listening for educators, students, and anyone curious about shaping the next generation’s skills in an AI-powered world. Whether you're at the start of your teaching journey or a seasoned professional grappling with assessment strategies, this discussion offers a fresh perspective on how to stay ahead. Prepare to rethink what it means to learn, teach, and succeed with AI as your co-creator, not your substitute.



    Get full access to My Data Guest at mydataguest.substack.com/subscribe
    Show More Show Less
    47 mins
  • Bonus Episode: Time Series Foundation Models, Agentic AI and Career Advice
    Jul 22 2026

    It’s always a pleasure talking to Luis and Joshua. This episode, we covered ground that ranged from practical career advice to the cutting edge of AI agents and time series modeling.

    What We Discussed

    Career Paths in AI: Luis and Joshua shared insights on navigating roles in machine learning and data science, from industry trajectories to the skills that actually matter when building real systems.

    AI Agents: We explored the state of autonomous agents. What’s hype, what’s working, and why the problems are harder than they initially appear. The conversation cut through the noise to focus on what’s actually deployable.

    Foundational Models for Time Series: The deep dive here was substantial. Time series modeling with large language models is a frontier area. Neither Luis nor Joshua pulled punches about the complexity. We discussed architectural choices, training considerations, and the gap between benchmark results and production performance.

    Why This Matters

    The intersection of general-purpose foundation models and domain-specific challenges like time series is where a lot of real innovation happens. If you’re building forecasting systems or working with temporal data at scale, the patterns we discussed apply directly.

    Thanks to Luis and Joshua for diving deep. This is the work that doesn’t make it into papers. It’s what practitioners actually need.



    Get full access to My Data Guest at mydataguest.substack.com/subscribe
    Show More Show Less
    1 hr and 25 mins
  • Ep. 19 - Web Knowledge for AI Agents
    Jul 8 2026

    For the latest episode of My Data Guest, I sat down with Antonio Malia, a dear friend since our Pisa days, now founder of Seltz, a startup rethinking web search from the ground up. Not for humans this time. For machines.

    The core problem

    Google, Bing, and every search engine we grew up with were built around a simple human loop: type a query, scan snippets, click a link, read the page. Antonio’s insight is that this entire interface breaks down for an LLM. A snippet is, in his words, a movie trailer. It hints at the answer but doesn’t give it. So the model either guesses or burns tokens and seconds fetching the full page. Multiply that by every agentic loop, and you get slow, expensive, unreliable agents.

    Seltz’s bet: build the crawler, index, retrieval, and ranking stack from scratch, optimized for a completely different reward function: one where the “user” is a language model, not a person.

    Why it matters: latency and trust

    Two things stood out from the conversation:

    * Latency compounds. If an agent can get a search result back in 100ms instead of a second, it can run ten iterations in the time a traditional engine takes for one. That’s the difference between a shallow answer and genuinely complex task completion.

    * Trust is the real target. Antonio framed it sharply: AGI, to him, is the point where we trust a machine’s completed process more than we’d trust another person’s. Web search is just the first, most obvious place models need reliable, constantly-updated access to the world. No more “airplane mode” after training cutoff.

    Full stack, small team, Rust

    Seltz owns everything: crawler, data connectors, knowledge pipeline, ranking. Building in Rust was a pragmatic call: efficiency, memory safety, and (notably) a much larger, more excited talent pool than C++ offers today. The team, partly ex-Amazon, partly IR-research veterans, is explicitly modeled on the “flat, high-trust, senior-heavy” structure Antonio admired at places like Netflix.

    On coding agents

    Antonio’s take on AI coding tools was refreshingly unfussy: they’re a boost only if you already know exactly what needs to be done. Used to offload thinking, they slow you down. Used to execute a decision you’ve already made, and to be challenged when they propose something better, they’re a genuine multiplier. Design decisions stay human.

    What’s next

    The next six months are about making Seltz’s product the main driver of value for customers who integrate it, validating the underlying research day by day rather than chasing a big splashy release. If you want to try it yourself: sign up in the console and connect via API/MCP, or just book a call with the team.



    Get full access to My Data Guest at mydataguest.substack.com/subscribe
    Show More Show Less
    37 mins
  • Ep. 18 - How AI Is Reshaping Careers
    Jun 17 2026

    Hiring in tech has always moved fast, but few corners of it are shifting as quickly right now as recruitment itself. In this episode, I sat down with Bruno Lomardo — “Loma” to most people — who leads talent acquisition at Rasa, the conversational AI platform. We talked about what generative AI is doing to hiring, the strange new problem of deep fake candidates, and why getting diversity right isn’t a box-ticking exercise but a competitive advantage.

    From recruiter to AI talent lead

    Loma’s path into tech recruitment wasn’t a straight line, and that turns out to be part of why he sees the field clearly. He’s lived through hiring booms and the cooldowns that follow, watching the market swing from “we’ll take anyone who can write a for-loop” to far pickier, signal-hungry hiring. That cyclical experience shapes how he reads the current moment — because the latest swing isn’t really about headcount. It’s about the tools.

    “We’ve been working with conversational AI for years,” he points out, and that long runway matters. At Rasa, AI isn’t a buzzword bolted onto the hiring process; it’s the product. So when generative AI started reshaping how candidates apply and how teams screen, the team had a head start in understanding both the upside and the failure modes.

    What generative AI actually changed

    The most visible effect of generative AI on recruitment is volume. Applications are easier to produce than ever, which sounds great until you realize a polished cover letter no longer tells you much. When everyone can generate a flawless, perfectly tailored application in thirty seconds, the signal that used to come from effort and craft starts to disappear.

    That creates a real discrepancy in the market. On paper, candidates look stronger and more uniform than ever. In conversation, the gaps show. Loma’s takeaway is that the screening burden has shifted — away from filtering for basic competence on paper and toward verifying that the person behind the application is real, capable, and who they say they are.

    Which leads to the most unsettling part of the conversation.

    The deep fake problem

    “Deep fakes are getting better and harder to detect,” Loma says, and he means it literally. Recruiters are now encountering candidates whose video presence, voice, or even live interview behavior may be synthetically generated or assisted. What started as an edge-case curiosity has become a credibility problem teams have to actively manage.

    The detection methods are evolving alongside the threat — verification steps, live and unscripted interactions, and tooling built specifically to flag manipulated media. But Loma is candid that this is an arms race, and the honest position is to assume the fakes will keep improving. The practical defense is process: build interviews that are hard to fake your way through, and don’t outsource your judgment entirely to a tool that can be gamed.

    Doing diversity right

    Where the conversation turned genuinely energizing was on diversity, equity, and inclusion. Loma’s framing cuts through the usual debate: this isn’t charity and it isn’t optics. “Doing diversity right creates better teams,” he says — and the emphasis lands on right.

    Done badly, diversity efforts become quotas chasing optics, which helps no one and breeds resentment. Done well, they widen the pipeline, surface candidates conventional processes overlook, and produce teams that make better decisions because they bring more perspectives to the table. The work is in the how: structured interviews, debiased job descriptions, broader sourcing, and consistent evaluation criteria that give every candidate a fair read.

    The road ahead

    Looking forward, Loma sees AI playing an even larger role in hiring — handling the repetitive front end, surfacing strong candidates faster, and freeing recruiters to do the human work that actually predicts success on a team. But the same technology that makes hiring more efficient is also what makes it harder to trust. The recruiters who thrive will be the ones who treat AI as a tool to sharpen judgment, not replace it.

    If there’s one thread running through the whole conversation, it’s that: the fundamentals of good hiring — verification, fairness, and genuine human evaluation — matter more now, not less.



    Get full access to My Data Guest at mydataguest.substack.com/subscribe
    Show More Show Less
    43 mins