MLOps World | GenAI Summit

TMLS 2026 has officially begun. Virtual Day is live and online right now: a full day of applied AI talks, open to everyone. Ten years in, the question is less "what can AI do" and more "what actually holds up." Let's get into it.

Foundation models work for NLP and vision. Applying the same paradigm to structured, sequential event data (transactions, interactions, behavioural signals) introduces challenges existing literature largely overlooks. We’ll discuss. Lin Liu, @Wealthsimple

From promise to practice: Dr. Mamdani's team at Unity Health Toronto deployed dozens of AI solutions in production to improve patient outcomes and hospital efficiency.

Scaling LLMs is fundamentally a systems problem. Not a hardware problem. Not a modeling problem. A systems problem, and that changes what you debug first when moving from prototype to production. Deepkamal Gill, The Vanguard Group. TMLS 2026 → torontomachinelearning.com/

@netflix 's Globalization team removes language barriers across 300+ million members in 190+ countries and 30+ languages. @cwolferesearch will walk through how his team is evaluating show synopses with LLM-as-a-Judge at TMLS 2026 → torontomachinelearning.com/

Your agent eval says accuracy improved. But did latency spike? Does your LLM-based metric even agree with human judgment? And is that 5% gain real or noise? Abhimanyu Anand, @elastic TMLS 2026 →torontomachinelearning.com/

An agent has a flow, and getting the flow right is critical. You can trust the result only if the path the agent took aligns with your architectural intent. Agents are processes. Michael Havey, OpsGuru. TMLS 2026 →torontomachinelearning.com/

Most large AI orgs have a data problem that isn't the one they think. It isn't missing data. It's data that exists, was built deliberately, and still isn't being reused. It sits in a pipeline nobody else can find. @Mendelsohn Chan, @Alation

For long-context workloads, the bottleneck is no longer raw compute. Memory, bandwidth, and latency become the real limits once sequences get long, and the assumptions that work for shorter workloads stop carrying over. Mehdi Rezagholizadeh, @AMD.

Semantic similarity is not the same thing as answer retrieval. Lean on embeddings as the default for every use case and you get systems that sound convincing while returning weak, incomplete, or confidently incorrect answers. David vonThenen, @NetApp .

Every software company claims to be becoming an AI company. Most are re-running the wrong playbook: treating AI like an infrastructure migration instead of a shift in how products are designed, shipped, and operated. Alet Blanken, @Workday . TMLS 2026.

The default assumption in AI security is that encryption costs performance. Fully Homomorphic Encryption, especially. The talk analyzes trade-offs between encryption overhead and latency, using open-source FHE and model optimizations. Tyson Macaulay, @01quantuminc

A robot that can recall what it did last week needs memory across three dimensions: spatial, descriptive, and visual. That is the architecture behind experience recall for temporal question answering in agentic robots. @stevewaslander , @UofT

Agent vulnerability is primarily architectural, not a model alignment problem. Fixing the model without addressing orchestration logic leaves the most exploitable attack surface untouched. Naga Sujitha Vummaneni, @Ripple TMLS 2026 → torontomachinelearning.com/

$26/month infrastructure. 200+ languages served. A core team of three. That's the Multilingual Climate Chatbot, a production RAG system, open-source and easy to adopt. Luis Ticas + Helena Yu, Sprout Climate

Real data can be a legitimate option for pre-training tabular foundation models - despite being underutilized in favour of synthetic data. It captures complex signals critical for downstream generalization. Anthony Caterini @Layer6AI is exploring that here at TMLS 2026