Fireside Chat: How Healthcare and Biotech Teams Build Secure, Compliant AI Infrastructure
Leaders from Artera AI and Union.ai came together to discuss how top healthcare and biotech teams build scalable, compliant AI infrastructure. Learn best practices in workflow orchestration in healthcare, data locality, security, and global AI deployment.
This transcript has been modified for reading. Listen to the recording for the full interview!
If you have questions about implementing reproducible workflows in your organization, schedule time with one of our AI Engineers.
<div class="button-group is-center"><a class="button" href="/consultation">Book A Meeting</a></div>
Host: Sage Elliott, Developer Advocate at Union.ai
Guests
- Ketan Umare, CEO & Co-founder of Union.ai
- Nathan Silberman, CTO of Artera AI
- Reda Oulbacha, Senior ML Engineer at Artera AI
Panelist Introductions
Sage Elliott:
Let’s introduce today’s guests:
Ketan Umare is the CEO and Co-founder of Union.ai, creators of the open-source orchestration platform Flyte. Ketan’s background spans cloud infrastructure and AI/ML platforms, with experience at Amazon, Citadel, Oracle Cloud, and Lyft, where he led the core ML team.
Nathan Silberman is the CTO of Artera AI, where his team develops AI-powered diagnostic tests to help personalize cancer therapy. Nathan holds a PhD from NYU and previously worked at Google on TensorFlow and startups like Butterfly and PathAI.
Reda Oulbacha is a Senior Machine Learning Engineer at Artera, where he develops the infrastructure to train and serve AI-based prognostic tests. Prior to Artera, he worked on ML systems at BusPatrol and Famed Data.
What is Union.ai and Why Biotech Teams Use It
Ketan Umare:
Thanks for having me! At Union, we saw a big gap between how traditional software and AI software are built. Traditional systems are linear, while AI development is cyclical and experimental.
Union was created to help teams build reliable, scalable, and fast AI infrastructure. In biotech, this is critical. Our platform helps accelerate breakthroughs in drug discovery and precision diagnostics by handling orchestration, compute, and compliance, so companies can focus on science.
One of our partners is Artera. We've been working with them for over a year and are proud to support their AI-powered cancer diagnostics across their entire stack, from research to production.
How Artera AI Uses AI & ML to Personalize Cancer Treatment
Nathan Silberman:
Artera’s mission is to personalize clinical decision-making in cancer therapy. After a cancer diagnosis, clinicians must determine the right treatment. More aggressive cancers need stronger therapies, but the risk level isn't always clear.
Our AI-powered diagnostic tests help clinicians by assessing:
- Risk – e.g., likelihood of metastasis
- Optimal therapy – based on the cancer’s biological profile
These tests use AI models to analyze gigapixel-scale histopathology slides, the same ones pathologists review under a microscope. The AI interprets biological patterns to inform clinicians on the best course of action.
Reda Oulbacha:
Exactly. These images are huge,100,000 by 100,000 pixels, and processing them manually is infeasible. AI enables large-scale parallel analysis in the cloud, and workflow orchestration makes that scalable.
Why Workflow Orchestration Was a Key Architectural Decision
Reda Oulbacha:
Artera’s constraints included:
- Rapid development of many AI-based diagnostic tests
- Consistent deployment across global regions
- Complex, multi-step workflows per test
A single product can involve multiple AI models and image preprocessing steps. So we needed a workflow orchestration system to automate and manage this complexity.
After exploring options, we chose Union’s Flyte for its:
- Developer velocity
- Global reproducibility
- Compliance features
Nathan Silberman:
Cancer care isn't one-size-fits-all. Every subtype needs a unique model. We needed infrastructure that could scale across cancer types, countries, and regulations, all while ensuring model versioning, monitoring, and reproducibility.
Before and After: Workflow Orchestration in Practice
Reda Oulbacha:
Before adopting orchestration, we ran our pipelines in single containers, validating one cohort could take days. With orchestration, we broke workflows into parallelizable steps. What took days now takes hours.
Ketan Umare:
And it’s not just speed, it's also about reliability, cost efficiency, and sleep. AI in healthcare can’t afford inconsistencies. You need traceability, version control, and auditability. Orchestration brings discipline.
Nathan Silberman:
For regulated models, we also need backtesting and simulation before deployment. You can’t just “ship and see”, lives are at stake. With Flyte, we can validate workflows reproducibly across environments.
What Is Data Locality and Why It Matters in Healthcare
Reda Oulbacha:
Let’s say you expand from the US to Australia. Your clinical contract might require data and computation to remain in-region, that’s data locality.
We achieved this in 4 steps:
- Storage: Patient data must stay in-region
- Compute: Servers must reside in-region
- Transit: Data must never leave the region during processing
- Artifacts: Intermediate outputs (e.g., logs, model features) must also stay in-region
Union’s global control plane lets us manage multi-region deployments from a single command point, without spinning up new teams per country.
Ketan Umare:
Exactly. Union separates control and data planes. Your workflows stay local, while your management is centralized. This is key for compliance, especially with multi-cloud or multi-region deployments.
Security & Compliance Considerations You Might Overlook
Reda Oulbacha:
Security works only if everyone in the org considers it their job. A few overlooked areas include:
- RBAC (Role-Based Access Control): Prevents data from crossing regional boundaries
- Data transit configuration: Ensure nothing “accidentally” routes outside the region
- Artifacts & logs: These may contain PHI and must be restricted
- Dependency chain vulnerabilities: Even GitHub Actions can leak secrets
- Secret rotation: Regularly refresh API keys and passwords
Nathan Silberman:
One subtle but vital point, training and validation data access should be partitioned. Regulatory bodies want to see that validation sets aren’t leaked or reused. We enforce this using RBAC all the way from S3 buckets to Flyte workflows.
Hiring & Closing Thoughts
Sage Elliott:
Before we wrap, Artera is hiring! Check their site for roles in:
- Full-stack development (React, Python)
- ML infrastructure
- Data platforms (handling gigapixel image ingestion at scale)
And Union.ai is also hiring folks to help build this infrastructure. You can find opportunities at union.ai or join the community via Slack at flyte.org.
Final Takeaways from the Panel
Reda Oulbacha:
Adopt orchestration earlier than you think. It saves time, improves reliability, and lets your team move faster in the long run.
Nathan Silberman:
Healthcare isn’t a “move fast and break things” domain, but you still need to move fast safely. Early investments in infrastructure and automation are what enable high-velocity, compliant innovation.
Ketan Umare:
You don’t have to build everything yourself. Partner with platforms like Union that let your team focus on what matters, science, not infrastructure.
<div class="button-group is-center"><a class="button" href="/consultation">Book A Meeting</a></div>
