<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Marco Saetta on Medium]]></title>
        <description><![CDATA[Stories by Marco Saetta on Medium]]></description>
        <link>https://medium.com/@marco.saetta?source=rss-62160a7c131d------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*XJmRkh6PQrFq0JC-jmqHuQ.jpeg</url>
            <title>Stories by Marco Saetta on Medium</title>
            <link>https://medium.com/@marco.saetta?source=rss-62160a7c131d------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 20:30:03 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@marco.saetta/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[From one AI product to an AI factory]]></title>
            <link>https://medium.com/doctolib/from-one-ai-product-to-an-ai-factory-e6337ffa75d9?source=rss-62160a7c131d------2</link>
            <guid isPermaLink="false">https://medium.com/p/e6337ffa75d9</guid>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-platform]]></category>
            <dc:creator><![CDATA[Marco Saetta]]></dc:creator>
            <pubDate>Thu, 09 Jul 2026 16:52:27 GMT</pubDate>
            <atom:updated>2026-10-24T15:26:11.970Z</atom:updated>
            <content:encoded><![CDATA[<h4>What two years as ML Platform PM at Doctolib taught me about scaling AI products</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*opAY3kzyb6YN-ggDPf40NQ.png" /></figure><p><em>Every company is racing to put AI into its products. Far fewer are solving the harder problem underneath: how do you make building good AI products something every team can do reliably, safely, and at scale? The platform is a big part of the answer, and it’s what I’ve spent the last two years on at Doctolib, as Product Manager for the ML Platform. This is an honest look back at that journey, from our first AI-native product to the foundations of an “AI factory”. I’ll cover what worked, what didn’t, and the lessons I think transfer to any company trying to scale AI.</em></p><p><strong>Doctolib</strong> exists for two missions: (1) to improve the daily life of health professionals, and (2) to help people live healthier. They can sound abstract until you see the scale that makes them real: roughly 100M people and 580k health professionals across several European countries rely on Doctolib, built by about 3,000 of us. Moving fast is part of the culture; doing it in healthcare, under real constraints on privacy, safety and reliability, is what makes it hard.</p><p>So what does a <strong>platform</strong> team actually do? Our job is leverage. A platform is the foundation of self-service tools, services, knowledge and support that lets feature teams build reliable products more efficiently, by standardizing the underlying building blocks so each team doesn’t reinvent them. Crucially, platform customers are internal: the dozens of feature teams building the products that patients and healthcare professionals actually use.</p><p>The <strong>ML Platform</strong> is the slice of the platform supporting the feature teams building <em>AI products</em>: applications whose core capability is driven by AI, such as the <em>Consultation Assistant</em> (an ambient scribe that drafts and structures doctors’ consultation notes), the <em>Phone Assistant</em> (which answers patients’ calls to handle booking and more), or the <em>AI Health Assistant</em> (a medical chat answering patients’ health questions and simplifying their interactions with the Doctolib app). Our scope is everything needed to <em>train, evaluate and serve AI</em>. We believe AI can push both Doctolib’s missions further, which is why one of our ambitions is to become a leading AI healthcare company. And you don’t get there by adding AI to a handful of products, but by making it something every team can build, reliably and at scale.</p><p>People sometimes ask what a <strong>Product Manager</strong> even does. The way I put it when friends ask: my job is to make sure the team’s effort goes where it creates the most impact for Doctolib. That’s it. Everything else is detail.</p><p>I joined right before the launch of the Consultation Assistant, Doctolib’s first AI-native product. What follows are the five transformations we’ve gone through, or are still going through, on the way toward becoming a leading AI healthcare company, each told through where we started, how we did it, what’s next, and what we learned.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Oq1xE9uhZ8dFgj2JGFeIMw.png" /></figure><h3>1. From fragmented to unified platforms</h3><p>A fragmented platform creates inefficiencies and inconsistencies. The first transformation, partially still ongoing, is to unify Doctolib’s separate platforms, because you can’t deliver AI features at scale on fragmented foundations.</p><h4>Where we started</h4><p>We had separate platforms side by side (Data, ML, and Engineering), each serving mainly one population and one kind of work: Analytics Engineers doing analytics, Data Scientists building and serving AI models, software engineers shipping “classic” product features. In practice, the ML Platform served only Data Scientists, the only role then in charge of AI backends.</p><p>It worked, but it came with real limits. <em>For users</em>, it meant fragmented journeys and an inconsistent developer experience across these platforms: painful access to data for training and evaluation, and inefficient transformation pipelines. At the same time, the nature of the work was changing: as AI products turned agentic (full systems with memory, workflows and tool calls, not just models), building them increasingly required software engineers, a population the ML Platform didn’t yet serve.</p><p><em>For the platform teams</em>, it meant overlapping scope, the cognitive load of maintaining redundant, non-differentiating plumbing, and infrastructure costs that were hard to justify. Acceptable at the time, but not if we wanted to really accelerate on AI products.</p><h4>How we did it</h4><p>The first was data, the fuel for any AI product: we had to make it accessible for training and evaluation <em>at scale</em>, without compromising security or privacy. After a cloud migration and more than a year of work, we merged the Data and ML platforms into a single <strong>Data &amp; AI Platform</strong>, a decentralized and fully governed architecture covering the full chain from ingestion and transformation to training, evaluation, and serving (Alexandre went deeper in <a href="https://medium.com/doctolib/building-a-unified-healthcare-data-platform-architecture-2bed2aaaf437">this article</a>). It now lets teams find and use the data they need, and freed my team to recenter on its real job: training, evaluating and serving AI.</p><p>But the merge itself wasn’t the hard part. What didn’t work was assuming the platform would be enough: hand people a unified platform and they don’t automatically change how they work. We’re still adapting on every side: the platform to new users’ habits, users to the new paradigms we push (an Analytics Engineer and a Data Scientist don’t build a data pipeline the same way), and platform teams to one another, defining the interface contracts that keep end-to-end user journeys consistent across teams.</p><p>The second challenge was serving those new builders. A platform built for data scientists suddenly had to give software engineers a consistent experience across both the Data &amp; AI Platform and the Engineering Platform.</p><p>What worked was a consistent aim: giving users the same experience by taking the best of both worlds (traditional software and AI products), rather than reinventing on each side. On deployments, for example, the ML Platform aligned to the Engineering Platform’s approach instead of rolling its own. And wherever a capability could genuinely serve both worlds, we built it once; feature switches and A/B testing were the first examples. The payoff is a more consistent experience for users and much less overlap between platform teams. Two deliberate moves pushed this forward: we reorganized the Data &amp; AI Platform under the Platform domain, and set shared OKRs between the Engineering Platform and Data &amp; AI Platform teams.</p><h4>What’s next</h4><p>Right now we’re rethinking the golden paths that cut across Data and AI products (ML training, evaluation, and context-engineering pipelines), where the interdependencies are strongest. The broader step is to continue harmonizing the Data &amp; AI Platform with the rest of the Engineering Platform: reusing shared infrastructure and unifying the experience across both. We’re confident this will deliver compliance-by-design, higher velocity, and more standardization when building AI products, moving us closer to our AI factory vision.</p><p><strong><em>What we learned:</em></strong><em> unifying platforms is an organizational problem more than a technical one. Tooling and infrastructure are the easy part; getting people (platform teams and developers) to change how they work is the real challenge.</em></p><h3>2. From technical enabler to platform as a product</h3><p>A platform can be tooling and plumbing feature teams merely tolerate, or a product they actively choose. This second transformation had little to do with technology and everything to do with how we work: treating the people who build on us as customers, not stakeholders. It’s what turns a platform that <em>exists</em> into one that teams actually <em>adopt</em>, and adoption is what multiplies the number of reliable AI products Doctolib can ship more efficiently.</p><h4>Where we started</h4><p>The platform was a technical enabler: useful, but run like infrastructure, not like a product. Most of the rituals you’d expect around a product were thin: little structured discovery, not much systematic user feedback, an OKR process that existed but was still messy. On prioritization, most of ML Platform’s effort went to unblocking the one or two feature teams building Doctolib’s first AI-native products. That made sense when we effectively had few customers; it didn’t scale to many.</p><h4>How we did it</h4><p>Becoming a product started with getting much closer to our users. All platform PMs now run user research interviews continuously; we co-build solutions and run proofs of concept with teams to get real feedback from the field. A habit I’ve come to rely on is living in the feature teams’ Slack channels. Half of it is just listening in: catching weak signals, seeing what a team shipped, noticing where they got stuck. That’s often where the best opportunities surface.</p><p>We also built a structured way to connect our work to its purpose. A platform product vision, refreshed roughly every six months, sets direction; user proximity feeds a backlog of opportunities, prioritized during discovery with RICE, adapted to how a platform creates value (productivity gains, reliability, avoided infrastructure costs, unblocked revenue); the top opportunities become quarterly OKRs, which break down into delivery epics. What worked here is the traceability: every Doctoliber (feature teams, management, platform engineers) can trace an initiative back to a real user problem, and know why it matters.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Xc5NGB_1D2cYQo-lf3R_uA.png" /></figure><p>The hardest part was prioritization itself. When you serve one or two teams, you just unblock them; when you serve many, you’re arbitrating between customers with real, competing needs. And a framework only takes you so far: this is a B2B-like setting where one feature team, or one user population (Data Scientists, software engineers…), can legitimately outweigh another, and a strategic bet sometimes overrides the score. RICE settles most calls; documented judgment handles the rest.</p><p>A second tension we simply live with: moving fast is in Doctolib’s DNA, so teams adopt platform solutions before they’re fully self-served or standardized. We cope with defined collaboration modes (X-as-a-Service, Facilitating, Collaboration), each with its own rules and level of support (Arthur describes them in <a href="https://medium.com/doctolib/our-path-to-implement-a-product-approach-to-data-4-5-a31c93992ee5">this article</a>). But it comes at a cost: heavy support load, less user autonomy, and divergence in how teams use the same solution, which limits how much one team’s work can be reused by another.</p><h4>What’s next</h4><p>Two priorities stand out. The first is metrics: moving from qualitative to quantitative. We’ve defined the core platform metrics in our product visions, but we don’t yet track all of them, or systematically. The next step is to track them consistently and tie them back to the opportunities they’re meant to move (productivity, cost, reliability, revenue), so the loop from user problem to measured impact finally closes. The second is self-service: pushing teams further toward autonomy, and leaning on AI agents to automate support as much as possible, something we’ve already begun.</p><p><strong><em>What we learned:</em></strong><em> prioritizing for many internal customers takes both a framework and judgment. The framework settles the easy majority, and documented judgment handles the calls it can’t. And user proximity starts a flywheel: more feedback brings faster improvements, faster improvements bring broader adoption, and adoption brings more feedback.</em></p><h3>3. <strong>From artisanal to platform-powered AI products</strong></h3><p>Every AI product we put in production is high quality, safe and reliable; in healthcare, it has to be. But the first ones got there in an artisanal way, optimizing only for quality, with slow iterations and a limited set of options: few foundation models and careful, hand-tuned prompt engineering. This third transformation is about industrializing that work through platform capabilities: making iterations faster, broadening the choice of models, optimizing cost and latency, and pushing quality past the point where prompting alone plateaus.</p><h4>Where we started</h4><p>Our AI products were first versions, or still under construction. A good foundation model and careful prompt engineering were enough to ship, and teams optimized for one thing: quality. Cost was mostly an ML Platform concern, along with much of reliability and incident management. Evaluation was artisanal: every product invented its own, with little shared tooling, thin observability, and subject-matter experts and PMs kept at the edges, working off non-integrated spreadsheets and fully dependent on data scientists for new experiments. Access to external models was manual and narrow, and our products ran almost entirely on generic foundation models, not our own. Each of the first three agentic products initially chose a different agentic framework. All of it shipped real, safe products, but it wasn’t yet a way to make them faster, better and cheaper.</p><h4>How we did it</h4><p>With limited platform capacity, we couldn’t tackle everything at once, so we sequenced deliberately: first give teams what they needed to <em>deliver and evaluate</em> quality AI products, then <em>start standardizing</em> how they build, and leave deeper optimization for later. Alongside that, the mindset shifted: feature teams now optimize for quality <em>and</em> cost, and own their systems in production (“you build it, you run it”). It played out through a few capabilities the ML Platform now provides:</p><ul><li><strong>Evaluation-driven development.</strong> Instead of every team hand-rolling its own checks, we provided a shared evaluation and observability platform. It turns AI development into something closer to a science, with a continuous feedback loop between experimentation and production. Teams have run more than 50,000 offline eval runs on it since early 2026, now at a pace of 2,500+ a week, and PMs and subject-matter experts contribute autonomously inside the loop rather than at the edges. This is what makes healthcare AI safe to ship, faster.</li><li><strong>Easier, faster access to more models.</strong> A single GenAI gateway gives every team governed access to a broad catalog of compliant models from different providers, swappable in minutes. More models and zero-effort to switch, along with efficient evaluation, mean more room to optimize for quality and cost and to iterate faster. It also brings resilience, with automatic fallback when a provider goes down.</li><li><strong>A common agentic framework.</strong> Agentic products now build on a shared foundation instead of each reinventing one, so teams start from a proven base rather than from scratch. Because everyone builds the same way, it also creates synergies across teams and lets us support them faster.</li></ul><p>We didn’t get everything right. We took too long to commit to a single agentic framework for the golden path, and while we deliberated, teams diverged, which cost us a painful migration later. Not deciding, it turns out, is itself a decision. The flip side is knowing when to change your mind: as the market moved, we reassessed and replaced tools we’d already bet on (our evaluation tooling among them) rather than defend the first pick. In a field where what took a developer a month can take a week a few months later, committing to a golden path <em>and</em> staying ready to pivot is a balance we manage constantly.</p><h4>What’s next</h4><p>The next optimization lever is our own models. So far, except for a few cases (such as a speech-to-text model we fine-tuned for medical language), our agentic products have relied on foundation models, context retrieval, prompt engineering and rigorous evaluation on Doctolib data. Now, as we start to seriously invest in AI research, we’re convinced that the next jump in quality will come from training on our own data. That data, built from millions of users, is the real durable advantage: no competitor has it, and most of its value is still untapped. Training on it means better quality and also potentially smaller, cheaper, faster models we can self-host, cutting cost and reducing our dependence on external providers, against a shifting geopolitical map for AI. One operational lesson of this transition is already clear: GPUs are scarce, so plan your capacity needs quarters ahead, and once you have them, optimize their use (e.g. smart queues, job concurrency).</p><p><strong><em>What we learned:</em></strong><em> with limited capacity, the order of investments matters as much as the investments themselves. Build the bricks that unblock the most value first, guided by how mature your internal ecosystem is: for us, evaluation and the GenAI gateway, then agentic framework standardization, then training on our own data.</em></p><h3>4. From platform capabilities to accelerators</h3><p>Having the platform building blocks isn’t the same as making them easy to assemble into customer-facing products, nor leveraging existing assets built by other feature teams on those building blocks. The fourth transformation moves from offering generic platform <em>capabilities</em> to building Doctolib-specific <em>accelerators</em> on top of them, so every team can go from idea to a running AI product in days, not quarters. This is our “AI factory” vision: many high-quality AI products across every team, not a handful built by specialists.</p><h4>Where we started</h4><p>The platform capabilities were under construction, and even when available, assembling them into a product was a manual job. Each team wired things together its own way, spending time on plumbing instead of focusing on business logic. Concretely, the first AI-native products took several quarters to reach production.</p><h4>How we did it</h4><p>Now that the core platform capabilities existed, we shifted to adding an abstraction layer on top of them, with a structured make-vs-buy strategy: buy the generic capabilities from best-in-class providers who ship weekly and stay state-of-the-art, and spend our own energy on the <em>accelerators</em>, the Doctolib-specific layer that assembles those capabilities into guided, reusable paths. Don’t build what others do better; differentiate where it counts.</p><p>The first accelerator we built is a golden-path agent template: built on the agentic framework we officially support, it comes with native integrations to our platform ecosystem (memory, observability, evaluation, model access, standard guardrails), turning a from-scratch build into a head start. The speed win is the clearest proof: what used to take several quarters now takes weeks. At a recent internal hackathon, a feature team went from nothing to an agentic product running in beta in production in about three weeks. We also inner-sourced it: feature teams contribute the features they need, the platform reviews contributions, keeps the standards and adjusts its roadmap, and every team benefits from what one team adds.</p><p>We’re not there yet, and the gap has two sides. First, reuse of <em>product</em> capabilities (the Doctolib-specific assets feature teams build on the platform, such as evaluation datasets, prompts, or trained models) is still low and largely manual. Second, our <em>platform</em> capabilities are sometimes used only at the surface, a bit like a new smartphone where most people tap a fraction of what it can do.</p><h4>What’s next</h4><p>We have several initiatives underway to close both sides of that gap. One is evolving the agent template into a more advanced agent builder: an interactive tool that, from a given use case, assembles the right existing product capabilities into a working system, wired together, with best practices like standardized evaluation on by default, so teams start from a contextualized base and focus on business logic and on iterating for quality and cost. Not everything can be reused from one product to the next, though: reuse comes in bands, from truly generic (borrow as-is) to genuinely one-off, so part of the job is spotting what a new product can reuse to move faster, and what it has to build itself. The intent is to keep building platform accelerators that promote reuse, ensure standardization, and abstract away technical complexity, until <em>every</em> feature team ships AI products quickly. That’s the path to the AI factory as we envision it:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*84kGdy0vHkh8673ZFWp7sw.png" /></figure><p><strong><em>What we learned:</em></strong><em> without shared standards and a bit of governance to make teams discover and adopt what exists, consistently with internal standards, even the best accelerator gets bypassed under delivery pressure.</em></p><h3>5. From manual coding to an AI-native SDLC</h3><p>AI has changed how software gets built, and Doctolib went from almost no internal use of AI coding assistants to nearly every engineer coding with one, day to day. That shift changed more than how engineers build AI products; it also changed how and why we expose platform capabilities.</p><h4>Where we started</h4><p>When I joined, in late 2024, AI for coding was still early even outside Doctolib; the tools that would drive mass adoption were only just emerging. Inside, the ML Platform used to provide <em>DoctoGPT</em>, our own version of ChatGPT to retrieve information from internal documentation, but nothing more.</p><h4>How we did it</h4><p>Adoption didn’t happen by accident, and it came in two phases.</p><p>The first phase was full adoption of coding assistants. It came from strong sponsorship at the top and a real willingness to experiment: a platform team fully dedicated to this “AI productivity” stream, full access to AI coding assistants for every Tech &amp; Product team, trainings, real-project experiments, an internal skill marketplace, and even adjustments to the career framework. By Q2 2026, about a year in, 100% of our engineers were actively working with coding assistants.</p><p>The second phase, in progress, goes beyond assistants, towards an AI-native software development lifecycle: remote agents delivering well-specified changes through partially automated workflows, with the human still in the loop. As trust is earned, the shift goes from human <strong>in</strong> the loop to human <strong>on</strong> the loop, where agents execute and engineers supervise, with risk-based automated reviews.</p><p>That shift has two consequences:</p><ul><li><strong>Platform teams are now building for agents, not just people.</strong> Every platform team now has to design for a new kind of user. We’re exposing our capabilities through interfaces agents can consume directly (MCP servers, CLIs and SDKs to interact with platform services, documentation, and higher-level abstractions through Claude skills and plugins), with the goal of making feature teams more efficient, more autonomous, and lightening our support load.</li><li><strong>AI for internal productivity and for customer-facing products are converging.</strong> AI coding was rolled out to all Doctolib engineers by the Engineering platform, a stream that ran in parallel to our AI products for patients and healthcare professionals, where the bar on quality and production-readiness is far higher. The two are now starting to meet. On one side, internal AI tooling can reuse the services and best practices we already built for AI products (the GenAI gateway, evaluation, observability) to raise the quality of internal work and rationalize costs. On the other, that heavy internal usage gives the ML Platform faster feedback, genuine learnings, and an additional solid stress test for its components, all of which flows back to strengthen the customer-facing AI products leveraging the same platform components.</li></ul><h4>What’s next</h4><p>We’re still early on both fronts, and what’s next is to keep pushing on each.</p><p>The first is building the platform for agents, and doing it judiciously. Using AI to build AI is powerful but easy to overspend on, so the discipline is to reuse what already exists before building our own (vendors increasingly ship agent-ready interfaces), to prefer deterministic code over burning tokens where plain code does the job, and not to over-engineer abstractions, since models are increasingly able to handle complexity themselves, so it sometimes pays to leave them room rather than wrap everything.</p><p>The second is the convergence between internal productivity and customer-facing AI. The balance to strike is how far to converge: internal productivity carries lower stakes, so we want shared standards without stamping out the freedom to experiment with new options where the constraints are lighter.</p><p><strong><em>What we learned:</em></strong><em> converging internal productivity and customer-facing AI is worth it, as the learnings compound. But keep them decoupled, so neither sits on the critical path of the other.</em></p><h3>Looking back</h3><p>Two years in, the through-line is clear: from supporting a single AI-native product to building the foundations for an AI factory, all to accelerate Doctolib’s two missions even further: improving the daily life of health professionals, and helping people live healthier.</p><p>One thing runs through all five transformations: none was a project with a neat deadline. They’re long-horizon shifts, advanced as background work, in constant tension with the day-to-day delivery we still had to keep unblocked. That tradeoff never fully goes away; the skill is in making it deliberately, not by accident.</p><p>We did a lot. I learned even more. And plenty is still unsolved. That’s exactly what makes the work interesting: the problems are real and hard, and I get to chase them alongside brilliant teammates who are deeply curious and doers.</p><p>If one of these transformations resonates, and you’d like to go deeper on it or share how you’ve tackled the same problems, I’d love to hear from you. You can reach out on <a href="https://www.linkedin.com/in/marcosaetta">LinkedIn</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e6337ffa75d9" width="1" height="1" alt=""><hr><p><a href="https://medium.com/doctolib/from-one-ai-product-to-an-ai-factory-e6337ffa75d9">From one AI product to an AI factory</a> was originally published in <a href="https://medium.com/doctolib">Doctolib</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>