Union.ai

On July 22, our Chief ML Engineer Niels Bantilan joins World Salon's fireside panel - "When the Pipeline Breaks: Building ML Infrastructure for Biotech R&D" - with ML and platform leaders from Character Bio, Octant, Karius, Health Advances, and Artera AI. Register below to hear experts talk about reproducibility, compute costs, and when a homegrown setup starts to break at scale. 👉 https://lnkd.in/dv-yZdsc #MLOps #Biotech #AIRuntime #Flyte

Seattle devs, MLEs, agent builders! Join me for an in-person hack night and build an agent loop from scratch. Niels Bantilan, Head of AI at Union.ai and creator of Pandera is visiting Seattle to get you started building your own Agnetic loops. We'll start with the core pattern (an LLM call, a tool step, and a feedback loop) and wire it into a Flyte 2 workflow you can actually run, trace, and rerun. Bring a laptop and an idea. We'll have starter code, tool integrations ready to go, and people around to help you get unstuck. Whether you've shipped agents before or you're wiring up your first tool call, we'd love to see you. Thanks to everyone who's already RSVPed, we've saved you a spot. 📅 RSVP here: https://luma.com/jjb6bjwc

When the Pipeline Breaks: Building ML Infrastructure for Biotech R&D Machine learning is now the core of how AI-native biotech does both discovery and diagnostics, running on pipelines that mostly began as homegrown scripts. As the work scales, that setup starts to break: reproducibility slips, compute costs climb, and the regulators arrive. The pressure lands differently across the field, on the research side, ML results that can't be reproduced and FDA guidance making model governance a price of entry; on the diagnostics side, HIPAA and GDPR shaping where data can live, and a clinical result a doctor acts on that has to be right every time. This roundtable brings together ML, data, platform, and diagnostics leaders from AI-native biotech to work through where pipelines actually break and what it takes to keep results reproducible, affordable, and regulator-ready. Speakers: Jasmin Bharadiya, PhD – Senior Data Engineer, Character Biosciences Joshua Paul – Head of Compute, Data Science, Software, and Operations, Octant Niels Bantilan – Chief Machine Learning Engineer, Union.ai Sivan Bercovici – Chief Technology and Business Officer, Karius Vivek Mittal – Partner and Managing Director, Head of Biopharma, Health Advances Nathan Silberman – Chief Technology Officer, Artera AI Event Details: Date: July 22, 2026 Time: 2:00 – 3:00 PM ET Location: Online Register: https://lnkd.in/dv-yZdsc

Most agents use tools by emitting one structured tool call at a time. The model picks a tool, waits for the result, picks the next, and every intermediate result flows back through its context. As the number of tools and steps grows, that gets expensive, slow, and brittle. code: https://lnkd.in/eTccvF7k Flyte: https://lnkd.in/g6sdBDdv Slack: https://slack.flyte.org/ Connect with Sage Elliott: https://lnkd.in/gt2DhW3W Code mode flips this. Instead of calling tools one at a time, the agent writes code that calls them as functions, runs that code in a sandbox, and returns only what matters back to the model. Why this works well in practice: models are trained on far more code than tool-call traces, so they compose operations more reliably in a language they already know. Loops, branching, filtering, and data transforms happen in the sandbox instead of as separate round-trips. Large intermediate results stay out of the context window, and tool definitions load on demand instead of all up front. The result is an agent that does more per step and spends fewer tokens doing it.

Batch inference looks simple until you actually try to make it fast and cheap. Then it's idle GPUs, cold starts on every new pod, stragglers holding up a whole round, and a dataset you can't fit in memory. I wrote up a sample architecture for running inference on 100M samples with near-maximal GPU utilization. Two of them, actually, and the difference between them comes down to one question: when a sample gets rejected and needs another pass, does it go back through the scheduler, or does the pod hang onto it and retry in place? Which one you want depends on your approval rate, and I've tried to be specific about where the crossover is. There's no new DSL here. It's Flyte primitives that already exist: reusable containers so the model stays hot, dynamic batching, streaming data -- put together in a way that keeps the GPU as the bottleneck instead of the scheduler. It gets deep into the specifics. If you're doing evals at scale or synthetic data generation or multi-round LLM judging, it's probably worth your time. Link to the blog in the comments.

Supervised fine-tuning teaches a model to imitate examples. Reinforcement fine-tuning teaches it to optimize an objective you define, which is how recent open models picked up reasoning, reliable tool use, and consistent output formats. GRPO (Group Relative Policy Optimization) is the method behind a lot of that work, and it is simpler and cheaper to run than the PPO setups that came before it.

Biotech and life sciences teams are doing some of the most important, compute-intensive work in the world (like protein folding, drug screening, genomic pipelines) and they've been forced to spend as much time on infrastructure as on science. Here's how the best teams offload infra to an AI runtime: https://lnkd.in/gjeBnFMz A few things an AI runtime does for these teams: AlphaFold, RDKit, GATK, legacy C++, in one workflow. Every task declares its own container image and resource profile. No shared environments. No version conflicts. No rewrites. Faster DBTL cycles. Flyte 2 caches task outputs by hashing inputs and code, so a tweaked parameter only reruns what actually changed. You never re-fold a sequence you've already scored, or re-dock a compound you've already screened. Full traceability. Union.ai versions every workflow, container, and artifact automatically. When a lab result contradicts a model prediction six months later, you can trace exactly what produced it. Teams like Delve Bio, Artera, Cradle, Rezo, Freenome, and LatchBio are already building on Union.ai. Pure Python. No YAML. No DSLs. Just science at scale. 🔗 https://lnkd.in/gjeBnFMz

Most AI teams think their security responsibility ends when data leaves their cloud. The Mercor breach proved otherwise. In the June edition of The AI Loop, we break down how AI has blown a hole in enterprise security that lets attackers walk in through the vendor entrance. We share the emergence of Zero Trust security architecture, which ensures no data transits outside your security perimeters onto vendor infra. Read the full piece here 👇 🔗 https://lnkd.in/eRggvx9x

This is what a real partnership looks like. Last week, ArteraAI's AI team was burning the midnight oil — running a massive number of simultaneous inference jobs, enough to make us reconfigure our cluster on the fly (and probably enough to single-handedly keep AWS in business 😅). The team at Union.ai didn't just watch from the sidelines. CEO Ketan U. and CTO Haytham Abuelfutuh were right there with us, hands-on-keyboard late into the night, making sure we crossed the finish line. Deadlines are stressful. Partners who show up make all the difference. Couldn't ask for better ones. 🙌