AGENDA 2026 - September 17
a sharp agenda to maximize qualitative connections
At dotConferences, we've designed conference agendas since 2012, with 45+ conferences held.
We heard our attendees' feedback: you want dense talks, an energetic rhythm, the ability to network and build new connections, and an overall fun and knowledge-packed day!
That is why:
-
The longest talks are 20 minutes. Yes, even the most deep tech ones.
-
Shortest lightning talks are after lunch.
-
You have breaks that are not too short, so you have time to network and see our partners.
-
You can ask your questions to the speakers live during interactive Q&A.
-
If your question wasn't selected, you can talk with speakers in the Speaker lounge after their talk, 1:1.
-
There is a single track, so you're not missing anything.
dotAI, is more than just one day, it's an experience that extends throughout the entire week, dedicated to AI builders through a series of partner events.
agenda
(*Might be subject of subtle last-minute changes)
⚡ are for 10-min lightning talks
9.00 am
Doors open, badge pick-up & check-in 🚪
9.00 am - 9.30 AM
Welcome coffee 🥐
Introduction
11.05 am - 11.35 AM
Morning break ☕
4.15 Pm - 4.45 PM
Afternoon break 🍰
Conclusion
6.40 Pm - 8.00 PM
Happy Hour! 🍹
join a side-event
Thursday
17 sept - 7 pm - 11 pm
Closing dotAI
w/ H Company
That's a wrap on dotAI, now let's have a drink 🥂
We're closing out dotAI the way it should be: everyone in the same room, music on, drinks in hand, and no rush to call it a night.
Come celebrate with the people who spent the day thinking about what's next in AI, or just the ones that don't want to miss out on a cool party 🤓
There'll be food, drinks, (dancing?), cool folks, and a few surprises.
Bring your team and whatever energy you've got left.
One last get-together, no agenda, just a good party!
Friday
18 sept.
dotjs
JS conf for JS builders
15+ speakers, check out the agenda now.
Combo available.
Come for dotAI, stay for dotJS.
Talks from Karen Li (GitHub), Nico Martin (Hugging Face), Lea Verou, Maximiliano Firtman...
For those who ship!
David Louapre
"Mechanistic Interpretability"
Mechanistic Interpretability is the field of ML that tries to open the black box and understand what is happening inside deep neural networks. Sometimes compared to the neuroscience of AI, mechanistic interpretability is of great interest for the security of AI models, but also to gain a deeper understanding of how they work and process knowledge through reasoning. Applied to LLM in the past few years, it has shown that language models represent concepts that can be identified, located and influenced, and it has lifted the veil on some of their internal mechanisms.
In this talk, I’ll introduce mechanistic interpretability in vision and language models, its recent results on LLM about concept representation and steering, and illustrate how it is possible to apply to open source models.
Guillaume Vernade
"Don't Become Your Agent's B****"
We were promised we’d become orchestra conductors sipping coffee while AI built our products. Instead, we became on-call terminal janitors doing hourly standups with hallucinating interns. Here is how to reclaim your sovereignty, kill DRY, and architect agent workflows that preserve human sanity.
Ziv Ilan
"When Prompting Runs Out: Post-Training an Open Model"
There is a specific moment when prompt engineering, retrieval, and better tooling stop paying and post-training starts. Most teams miss it in both directions.
This talk maps the cheaper levers first - a newer base checkpoint, context engineering, RAG, tools, constrained decoding, routing - and the exit condition for each. Then what comes after: choosing between SFT, DPO, RL with verifiable rewards, and distillation by the learning signal you actually have, with real data and compute needs per method.
We'll use Nemotron as the worked example, having the weights, the datasets, and the training recipes all public - you can read every step, then run it on your own domain.
Kevin Scaman
"AI for Prediction: Why You Need a Statistician, Not a Philosopher"
AI has become remarkably good at talking, reasoning, and generating content. But many of the decisions that matter most in business are not language problems, but prediction problems: Which customer will churn? How much will demand grow? Where is the risk? What happens next? These questions live in the world of structured, tabular data. And while Large Language Models were trained to understand and generate language, a new class of Large Tabular Models is being trained on billions of tables to learn the statistical patterns behind prediction. In this talk, I’ll explain why these two kinds of models have fundamentally different strengths. Think of them as complementary sides of an AI brain: one built to understand and communicate, the other to predict from data. Through a series of simple experiments, we’ll look inside the surprising limitations of LLMs when asked to make predictions and see why, when an important decision depends on your data, you may want a statistician rather than a philosopher.
Patrick Brosset
"WebMCP: Stop scraping the Web, start calling it"
Today, most AI on the web relies on a model visually interpreting pages or scraping HTML and then guessing which clicks to make.
WebMCP is the bridge from “AI can read pages” to “AI can safely and predictably do work on pages” by replacing guesswork with explicit tool contracts.
In this talk I’ll show how WebMCP turns multi-step, high-friction workflows into natural-language instructions that map to structured tool calls you can inspect. We’ll see how WebMCP eliminates DOM brittleness, is resilient to UI churn without shipping a separate server-side integration for every scenario and allows for a practical adoption path.
You’ll leave with a concrete mental model for what WebMCP is and how it can put users and web developers back into agentic browsing.
⚡ Xuan-Son Nguyen
"llama.cpp - faster and easier for everyone!"
llama.cpp started as a single-file experiment: run a large model on a laptop. Three years on, it's a full local-inference platform. We'll go under the hood of its biggest recent gain, speculative decoding: why it's "free," and how llama.cpp drafts tokens several ways: draft model, n-gram, MTP, and the newest additions, DFlash / DSpark. And since fast tokens are only half the story, we'll finish with a live demo: one command to install, running in seconds.
⚡ Guillaume Blaquiere
"From Blank Page to Production: A Year of Building AI Agents at Carrefour"
When we started deploying AI agents in early 2025, there were no frameworks or templates—just a blank page. Month after month, as tech stacks evolved, we integrated RAG, observability, and evaluation systems into our ecosystem. This journey was anything but a straight line, and hands-on experience completely reframed our vision of what agents can do. In this retrospective, I will share our key takeaways, unexpected challenges, and the ultimate Do's and Don'ts of agent development and deployment.
⚡ Greg Qualls
"Your AI agents have ADHD"
An algorithm figured out I had ADHD before any human did.
It was 2020. TikTok kept feeding me videos about how the adult ADHD brain works, and one after another, they described me.
Years later, I started building AI agents. And I watched them fail in the exact same ways my brain does. Not as a metaphor. As the same structural problem, close to one to one.
The part I didn't see coming: the wiring behind those failures is the same wiring behind the strengths, in a brain and in an agent alike. That changes what you're supposed to do about it.
People with ADHD have quietly solved this for decades. In ten minutes, I'll show you how to point those same tools at your agents, and you'll walk out with something you can use Monday morning. The lens comes from the last place anyone expected.
⚡ Aygalic Jara
"Hallucinations: Harvesting Uncertainty"
LLMs are startlingly confident even when they're completely wrong, and disclaimers, RAG, and tool use only take you so far. But there's a signal we mostly ignore: models carry a usable internal sense of when they're about to hallucinate, readable before generation starts and token by token as it proceeds.
This talk covers why hallucinations are structural rather than accidental, how uncertainty can be harvested from the model's internals, and what you can do with it: fine-grained hallucination detection, model and human escalation, abstention, and observability into the queries your system handles badly.
The goal: reframe model uncertainty as an engineering tool, not a research curiosity.
⚡ Leo Arsenin
"How Do We Compress Giant AI Models to Run at the Edge?"
Quantization is the go-to method for making LLMs smaller, but it's lossy, quality degrades in ways that are hard to predict and impossible to reverse. We found a different path by looking at how model weights are actually stored. In BF16 format, every weight has an 8-bit exponent, but only 16 values out of 256 cover 99% of all weights in a typical layer. Information theory says you need just 2.6 bits to represent that distribution, which means 5.4 bits per weight are pure waste. Unweight applies Huffman coding to the exponent stream, decompresses directly in fast on-chip shared memory, and feeds the reconstructed weights straight to the tensor cores, they never round-trip through slow main memory. The result is a 22% model size reduction that is 100% lossless with bit-exact outputs, saving roughly 3 GB of VRAM on Llama 3.1 8B alone. For Cloudflare, this means fitting more models on fewer GPUs across our 330+ city network, making inference cheaper and faster for everyone. We open-sourced the GPU kernels and published the full research paper, so you can use this technique on your own models today. If you care about running AI sustainably and affordably, this talk shows you a compression approach that doesn't ask you to sacrifice quality.
⚡ Hélène Philippe
"Radiological AGI: How Close Are We?"
AI in radiology has split between specialized point-tools that excel at narrow tasks but don't generalize, and vision-language models (VLMs) whose versatility and zero-shot capacity are undermined by hallucinations, imprecise measurements, and poor reproducibility. Our work examines how agentic systems could bridge this gap, pointing toward a path for building autonomous radiological assistants — and toward what real progress on Radiological AGI actually looks like.
Stanislas Polu
"Will the future of code be formal or informal?"
Human attention (in the form of reviewing agent-generated code) has become the bottleneck for modern software teams. As code becomes cheaper to write and more expensive to trust, will it remain the primary artifact developers work with? I’ll explore the spectrum from formal methods made practical by agents to futures where developers no longer work with code at all. I’ll argue that code is here to stay, but that neither fully informal nor fully formal specifications are desirable. The interesting future might lie between prose and proof: code augmented by local, structured natural-language contracts that humans can review and agents can audit.
Ian Massingham
"How Anthropic Uses Claude"
Every developer in the room can use the same models Anthropic does. So why do some teams turn them into measured results while others are still running demos? In this talk, you'll learn how radical new approaches to working with Claude have driven productivity at Anthropic as Ian walks through the five operating principles behind that, with the engineering, security and sales cases to copy into your own organizations.
Gaëtan Brison
"The Moving Target: Benchmarking Dynamic AI Systems End-to-End"
AI systems don't stand still — so why should our evaluation frameworks?
At Doctolib, we are building a complex ecosystem of AI agents spanning triage, clinical interview, retrieval, memory, booking, and image understanding. Evaluating each component in isolation is no longer enough. In a medical context, what matters is whether the system as a whole performs reliably — across multiple turns, multiple sessions, and across successive versions as the product evolves.
In this talk, we explore how to build an end-to-end evaluation framework for dynamic, multi-agent AI systems in high-stakes healthcare settings. We tackle the core challenge head-on: how do you test a moving target?
We introduce the concept of synthetic users — AI-powered personas that simulate realistic patient journeys, enabling systematic and reproducible evaluation at scale without compromising real patient data. We show how synthetic users allow us to stress-test agent pipelines across diverse scenarios, edge cases, and failure modes — including interactions with external systems.
But evaluation goes beyond task completion. In a medical setting, what the AI says matters as much as what it does. We dive into the challenge of managing context integrity across long multi-turn conversations — ensuring the system neither loses track of critical information nor hallucinates facts that were never grounded in the conversation. We present our approach to measuring groundedness: how well agent responses stay anchored to verified sources, retrieved documents, and the actual patient context — rather than drifting into confabulation.
Maxime langevin
"We Thought the Digital World Would Be Easy to Simulate. Was it?"
Self-driving cars stopped learning on real roads years ago. They learn in simulators, because a real crash is too slow, too dangerous, and too expensive to learn from at scale. Robotics went the same way. Digital agents, the software that clicks, types, and navigates your apps, are now making the same move, and it raises a puzzle.
You would expect the digital world to be the easy one to simulate. It is already digital: just run the real browser. The physical world, with its messy sensors and continuous physics, should be far harder. Yet simulation came first for cars and robots, and only now for software. Why?
The answer reframes what a world model actually is. Training an agent with reinforcement learning needs millions of resettable, parallel, rewardable trials, and real apps are too slow, flaky, and unobservable to provide them. Recent work like Qwen-AgentWorld shows the alternative: a language model that simulates the environment itself. It works not despite the digital world's complexity, but because that world is authored in symbols and text — accessibility trees, not pixels — which is exactly what language models already compress. This talk explains why the sim-to-real gap that torments robotics nearly disappears when the "world" is made of language, and what that means for how we build capable agents.
⚡ Diane Le Douaron
"How to Be Correct at Scale With No Ground Truth"
Most LLM work tolerates a near miss. Catalog transformation cannot afford it: expected outputs like fixed value lists and fixed rules demand a deterministic standard of correctness. This talk covers how we evaluate correctness when capabilities span several pipeline blocks and no ground truth exists.
Joffrey THOMAS & Henry Lagarde
"Own your agent"
As we transition from conversational chatbots to autonomous agents, relying on closed-source APIs has become a major bottleneck for production-grade applications. To build reliable agents that handle critical tasks, developers need absolute control over their models, their data, and their infrastructure.
In this talk, we will discover why Open Source is the only viable foundation for truly autonomous agents, offering the freedom to customize, the power to secure your data, and the independence to build without boundaries.
Christian Keller
"Beyond the Chatbox: What Changes When AI Runs Where You Are"
Most AI is built as if the answer is always a language model in a data center, reached through a text box. That breaks the moment your product runs on a laptop, a phone, or a pair of glasses. This talk starts from the constraint instead of the model: how form factor drives model choice, why specialized models solve problems no amount of prompting will, and why open weights make that design space navigable.