How the FDA Is Scaling Secure AI with Databricks for Government

Summary

Databricks on AWS GovCloud provides FedRAMP High and IL5-certified infrastructure for federal agencies working with export-controlled, defense, or highly sensitive controlled unclassified information, with capabilities including serverless compute, Unity Catalog, model serving, and Genie available today.
The FDA, acting as FedRAMP High sponsor, migrated from fragmented legacy center-based systems to a unified Databricks lakehouse called HALO, scaling from one regulatory center to six with 5,000 users and 8,000 jobs and zero-downtime migration.
The FDA's ELSA platform uses Databricks AI capabilities on GovCloud to assist drug review workflows, modernizing over 1,000 legacy dashboards into interactive, AI-powered experiences while eliminating technical debt.

How the FDA Is Scaling Secure AI with Databricks for Government

AI in the public sector has moved from pilot to mission critical, and the agencies leading the way are the ones that have mastered the path from experimentation to production in FedRAMP and IL5 environments.In this video, we will share the best practices customers should follow to get their AWS GovCloud environment ready for the most secure and mission-critical workloads — covering infrastructure-as-code security patterns, governed foundation model access, lineage and auditability, and the networking configurations needed to connect Unity Catalog to the right data sources. The U.S. Food and Drug Administration (FDA) will then showcase their experience getting their Databricks environment ready to support the most advanced workloads using Unity Catalog and the platform's GenAI capabilities. After migrating from legacy systems to Databricks on AWS GovCloud, FDA is now using serverless compute, model serving, Unity Catalog, and Genie to modernize over 1,000 legacy dashboards into interactive, AI-powered experiences — eliminating technical debt and accelerating mission impact.
Talk By: Bhagath Yarlagadda, Operations Research Analyst, FDA ; Filippo Seracini, Senior Staff Product Manager, Databricks ;

Chapters

FAQs

What is Databricks on AWS GovCloud and who needs it?

Databricks on AWS GovCloud is a FedRAMP High and IL5-certified deployment of the Databricks Data and AI platform designed for federal agencies handling regulated data such as export-controlled technology, controlled unclassified information, or data where unauthorized disclosure could lead to catastrophic events. It provides the same Databricks capabilities — including Unity Catalog, Genie, Lakeflow, and model serving — in a hardened, compliant infrastructure.

How did the FDA approach its migration to Databricks?

The FDA migrated from fragmented legacy systems across individual regulatory centers to a unified enterprise data platform called HALO, starting with a single center and scaling to six while maintaining zero downtime throughout. As the FedRAMP High sponsor for Databricks, the FDA played a key role in enabling the platform for the broader federal government.

What is the ELSA platform and how does it use Databricks AI?

ELSA is the FDA's AI-assisted drug review platform built on Databricks on AWS GovCloud, using serverless compute, model serving, Unity Catalog, and Genie to modernize legacy dashboards into interactive, AI-powered review experiences. It helps FDA reviewers process regulatory submissions more efficiently by surfacing relevant data and insights through natural language interfaces.

What Databricks capabilities are available in the GovCloud environment today?

Available capabilities in the Databricks GovCloud environment include Genie for natural language SQL, Unity Catalog for governance and lineage, Lakeflow for pipeline orchestration, model serving, and serverless compute. This video also covers a Terraform-based hardened infrastructure setup and a five-day onboarding path for agencies getting started.

Full transcript

[00:08] Good afternoon everyone. Thanks for being here today. Um in this talk we are going to work today we'll present on data bricks on AWS cloud and then more so we're going to focus and discuss on
[00:20] how the FDA using data bricks and asc cloud to scale their usage and their
[00:33] My name is Philippo Serachini. I'm a product manager in data bricks. I am responsible for all the compliance offering worldwide for data bricks as well as the regulated clouds and I'm here on stage with Bat Yard Lagada.
[00:46] >> Go ahead. >> Thanks Philippo. My name is Baty Lagarda. I'm with FDA's office of digital technology and I'm leading the infrastructure and platform to modernize and consolidate our enterprise data platform which is Halo.
[01:01] You probably already saw this line so I'm not going to spend much time on it. Just please don't take any word as a strong commitment is all all the statement we do a forward-looking statement. So in this talk we're going to discuss
[01:14] today first of all we're going to look at why go cloud who needs gocloud. We're also going to discuss what data bricks offers today in GC cloud and have a little bit of a roadmap chat and then
[01:26] I'll pass the mic to Bagger to show you how the FDA is leveraging uh data bricks on on Golf Cloud for their workloads. So let's get started.
[01:39] The first point who needs a GC cloud environment why Golf cloud lot more restricted a lot more regulated. So it's um It's an environment that allows you to
[01:51] process workloads that otherwise will not be possible. So if you have workloads for example that are protected under ITAR international traffic and armor regulation think of weapons, ammunitions, military
[02:05] vehicles or data and data sets protected by or regulated by export administration regulations er think of uh they usually um
[02:18] regulate dual use technologies technologies that are benign in nature but could also be exploited for military use, nuclear technologies, toxin chemicals,
[02:30] uh laser technologies, propulsion technologies, that kind of stuff. You also may be under fedum high or the DoD impact level five. So fedum high
[02:42] usually is a regulates uh control unclassified information which the disclosure of that could lead to catastrophic event and even loss of life. Here I have in this slide a few
[02:55] example like healthcare records or um financial system federal financial systems. The DoD is also even more restricted with the IL5 or for higher
[03:07] sensity CUI or national security system data. Lastly like you will hear later when uh during the FDA part also for example the FDA received submission for
[03:19] new drugs, new medications and all the data associated with it. that needs to be protected. Go cloud is an environment that allows you to do so. So if any of those workloads and any of
[03:31] those data sets sound familiar, go cloud is your answer. So data bricks on the list of cloud we launched about a year a year year and a half ago has been a rocket ship of a of an environment. We already have 70 plus customers most of
[03:44] them are in production today. 2,300 workloads and is the third fastest region in terms of revenue, fastest growing region that's revenue at data bricks all up across all commercial regions across Azure, AWS and GCP.
[04:00] It's comprised two environments. One we call it the community and the other one is the DoD. The only difference between the two environments or two shards as we call them is that is connected to the
[04:12] neetet network. If you're wondering whether you are connected to the different network then you are on the community chart and you should use the other other one. Okay. As I mentioned earlier pedum sorry data
[04:26] bricks on cloud is is certified or authorized for pedram high dial 5 hea high trust and supports ITAR and er workloads.
[04:44] Arabic on WSO cloud is a missionritical system available today. As a matter of fact is the most advanced data and AI platform you have on AWS cloud today. He has a tons of capabilities both on the
[04:56] data analytics side as well as the AI analytics side available today. Here I mentioned few of the la recent launches we had on AWS cloud in the last few
[05:08] months. So for example, now you can create your data uh data app on top of data bricks apps. You can run your analytics workloads on serverless SQL. You can invoke and have a conversation
[05:21] about your business data with data with Genie one. You heard today during the keynote as well as Genie agent and Genie code. You can also invoke frontier
[05:33] models through foundational model APIs or use vector search to index your data. Later this summer, we are going to expand our capabilities even further. Our objective is to have GC cloud as
[05:46] close as possible to the commercial side and we're going to bring all the feature that you heard today eventually will land on GC cloud. So later this summer, we're also going to bring default storage and we're going to have we're
[05:59] going to add the capability of sharing data to and from GC cloud to other regions. For example, you could have your workspaces on on the commercial side on the low side and share data to
[06:11] and from with your GC cloud environment regardless Azure, AWS or GCP. We're also going to bring a zerobust to simplify your caf cafka injection
[06:24] workloads. And from the geni side, we're going to bring zerobus. Sorry, that's for zerobus. And from the geni side, we're also going to bring batch inference.
[06:41] Later this year, you heard a lot about lakebase that is coming as well to go cloud as well as clean rooms to safely and uh share data with third parties in a very control and govern way. We're
[06:53] also going to expand on our AI portfolio bringing more model serving capabilities. Think of model OSS open source models and more frontier models. We're working to bring Gemini to toss
[07:06] cloud. The ability to invoke Gemini models from your environment on AWS cloud and all the AI function AI pars and so on and so forth. Those are going to come later this year.
[07:21] An environment regulated like the one in GC cloud to protect sensitive data from Fedrump to IL5 and so on requires a lot of hardening, a lot of controls, a lot of compliance control, a lot of requirements that need to be met.
[07:33] The good news is that databix has done a lot of work to bring you those controls building them into the platform. The even better news is that we have done a lot of work and we created this security
[07:47] reference architecture to basically tell you how to configure them. But even better, we actually create a Terapform script that you can run and apply the security reference architecture and create your workload, your workspace
[08:00] immediately with all the hardening in place. And the best thing is we built so many of these controls directly into the platform to abstract away the complexity of operating an environment like that.
[08:14] What that results is that you have actually the same security reference architecture both on golf cloud and on the commercial side which directly translate as a benefit for the users to
[08:26] have the same product experience the same product and the same experience that you have on commercial. So it's no compromised experience is no different product. is the same capabilities with no extra complexity
[08:39] and all is basically takes you is to run a a terapform script. Thanks to all this hardening that we baked into the infrastructure and the
[08:51] terraform script that I mentioned, you can literally go from not having an environment on gocloud to have a production ready audit ready environment in less than 14 days. Let me walk you through this. So on the first day you
[09:05] create your AWS GC cloud environment account I should say and you can create your data bricks account on AWS. To do that for example you could use the PGO
[09:17] marketplace environment the offering that we launched a few months ago. You then run the terapform apply that will create your workspace the first
[09:30] workspace with the secure baseline applied to it. Think of like the private link custom managed keys and the compliant security profile. If you don't know what the compliance security profile is, is basically a security baseline that we apply on your
[09:43] infrastructure that applies all the applicable controls to meet the regulatory controls. For example, from Fedrum High or IL5 or it's all apply immediately to you
[09:55] through the Terapform script. On the second day, you work on your single sign on configuration and you enable your Unity catalog to start importing your data and have it govern audited so that you can
[10:10] get ready for your regulated workloads. You then spend the rest of the week familiarizing yourself with all the capabilities. think like Genie or run your first workloads analytic workloads using serverless SQL and so on or maybe
[10:22] start creating the first dashboard with the IBI on on your govern data through Unity catalog on the second week you keep going and keep familiarize it more and you maybe
[10:34] by the end of the week you're ready to run or invoke the first frontier models through the foundation model APIs again all of this in an environment that is already configured to be ready to to go
[10:46] through a atto for fedum high and IO5. Obviously we cannot make you compliant but we we can help you creating environment as as harden as possible. So it's you have a lot less work to do to
[11:00] get it compliant. You could do obviously all of this by yourself but it would require a lot of months of work of hardening of bespoke hardening to achieve at best the same result. So you have a lot of these
[11:14] hardening, a lot of these controls built into the platform, easy for you to apply through a terapform script. And if you wonder how long it takes for this terapform script to run, here you
[11:27] have a screenshot of the running time, three and a half minutes. So once you create your AWS account and your data bricks account, you can run the Terapform
[11:39] apply. Three and a half minutes later, you will have your first workload, your first workspace deployed. You have a screenshot here for the Gulf cloud. That's the workspace with all the hardening and I mentioned
[11:52] again you will only thing you the only input you will have to provide is the configuration of your Terapform script with your parameters and so on. So from going in uh instead of spending
[12:04] weeks or months reading through documentation, figuring out how to harden, how how to configure, three and a half minutes, you have your first workloads up and running and you're good to go. But I don't want you to take my word for
[12:16] it. I mean obviously this is true but I want you to hear from Bart and so that he will walk you through on how the FDA and their using Golfcloud and Dix on
[12:28] Golfcloud and their journey on how to using that for to process their related workloads. Thank you.
[12:46] >> All right, good afternoon everyone. I hope everybody can hear me. Uh so I just want you to imagine steering an aircraft carrier while rebuilding its engine mid ocean. That's exactly how the enterprise data modernization works in federal
[12:58] regulatory agencies. The our mission cannot stop but the modernization must go on. So FDA is one of the most data inensive federal agencies in the world. Every
[13:10] drug approval, every food supplement determination, every device clearance, all of it depends on our ability to access the data quickly, securely, and responsibly. I want you to take take you
[13:22] through the journey on how we made it a reality. So, this is the number I want you to hold on to. One in three Americans are
[13:36] touched by a FDA decision today. Not in a year, today. That's the stakes behind every line of data we manage. FDA regulates 20 cents of every dollar
[13:48] of US consumer spending. We are more than 16,000 employees working across 200 offices and labs. And also we manage more than 300 product categories
[14:00] from prescription drugs to cosmetics to medical devices. With this volume, the data is not a back office function anymore. It's a public health imperative.
[14:12] When our systems are slowed, siloed, insecure, lives and outcomes are at stake. That's the context of everything I'm going to share with you today.
[14:28] The title of this slide says it all. We have moved from one regulatory center using the data platform to making it as enterprise data platform. FDA reviews millions of data points every year. Submissions from industry, clinical
[14:40] trials data, inspection findings, adverse event reports. The volume is immense and the complexity is also extraordinary. Our IT mission is straightforward. Deliver a secured, governed AI ready
[14:53] platform that allows FDI to innovate faster and serve the public better. In 2020, we have started our journey with data bricks using the legacy environment which we call PBC.
[15:06] And along the way, we've added the new capabilities that data bricks released. But not until 2025 when we moved to the GVC cloud high, we were really able to see the benefits of the Unity catalog framework or additional capabilities
[15:18] that the GVC cloud environment provided. And FDA has multiple centers, each focused on a different part of our regulatory mission. Drugs, devices,
[15:32] food, biologics, tobacco, and more. For a long time, each center ran their own data systems on their own. That worked up to a point, but it created silos,
[15:44] duplication, and also made hard for everyone to work together. So, we have adopted the multi-tenant architecture. In simple terms, all the centers share the same platform,
[15:56] but each one has its own secured govern space. Think of it like an apartment complex. Everyone shares a foundation infrastructure, but each center/tenant
[16:08] uh has their own building with their own lock. No one walks into someone else's space without their permission. And each center manages their own data and policies. with Unity catalog. Now it becomes very
[16:21] easy for us to share the data between these centers and I just want to show you the user adoption. Back in 2020 probably we have roughly around 500 users we started
[16:33] with. Now it's almost 6,000 more than 6,000 users who have been on our platform and we expect that counts to grow to more than 10,000 by 2028.
[16:50] All right. The previous slide represents where we are today. But every modernization starts with an honest look on where you are. Where we were is in a place of significant technical debt. So we had siloed data across FDA, many centers and programs. Different center
[17:04] help build their own pipelines, their own data stores, their own tools. This led to a lot of data redundancy, inconsistency and also duplicated effort. The result is delayed decision-m
[17:18] analysts, scientists spending too much time finding and reconciling data rather than deriving insights from it. Due to multiplatform environment, we had significant operational overhead as
[17:31] well. And also there was a widening gap between what our scientific and regulatory mission demanded and what our infrastructure could deliver. So that brings us to why we modernized.
[17:45] We needed faster insights. We needed enterprise data governance. We need to enable AI responsibly. We need a scalable future ready data foundation to support the agency's mission for years to come.
[18:03] And one strict rule that we all have to ad specifically at FDA is to the FISMA high and fedramp high baselines because we have public health data. We have industry trade secrets. So it's our duty to secure the data as much as we can.
[18:22] So this slide captures the ma three major milestones of our modernization journey and I want to spend a little extra time here because this is the really heart of the story. The first milestone is the Fed Ramp high authorization.
[18:34] This is an inflection point for FDA. FDA took action to sponsor the Fedramp high authorization. We partnered actively with data bricks to clear the path for unlocking the capabilities we
[18:46] knew we needed. Fedramp Fedramp high authorization gave us confidence to modernize at scale while fully meeting FDS security and
[18:58] compliance obligations. This was the prerequisite that unlocked everything else. The second milestone is the GC cloud migration. We performed a strategic replplatforming. We have moved over
[19:11] 5,000 users uh users and also migrated more than 8,000 jobs and pipelines with a zero downtime. I would say zero.
[19:25] Our scientists and analysts continued their work uninterrupted while we rebuilt the foundation underneath them. So this opened up for new capabilities like Unity catalog governance, serverless compute, model serving, schem integrations which made our life easier.
[19:41] The third milestone is the Unity catalog. This is the governance that we we we needed and we have refactored more than thousand data pipelines, more than 4,000
[19:53] notebooks as part of our migration. And with the Unity catalog, we got the end to end lineage and fine grain access control,
[20:10] the real stuff. So let me talk about what this actually has meant in practice because modernization is only meaningful if it delivers real results. As part of our enterprise data platform which we started last year in matter of
[20:23] few months we are onboarded eight centers and 30 programs across and also we have consolidated more than 40 data sources of application and
[20:35] submission data systems. Now our workforce is more empowered and also the collaboration across FDA is fast. The transparency has improved and our
[20:47] security posture is more robust. So how does this consolidation helped? That brings us to our AI enablement. We have integrated with Elsa, our own
[20:59] FDA's enterprise AI platform, enabling AI innovation at scale we couldn't achieve before. We now have AI use cases in production, in pilot, and more than 10 in flight.
[21:19] demonstrating that responsible AI is possible in the federal regulatory environment. Now coming to the operational side of it, we have gained 30% more than faster query response times using the SQL warehouses for the BI workloads. At the
[21:33] same time, we've reduced our compute cost by more than 20%. And we have reduced our the time we spend in provisioning the permissioning and also the data sharing by more than
[21:45] 75%. And we have reduced our operational overhead by more than 35%.
[22:00] All right. This brings us to the most exciting part. One of the most exciting outcomes we have achieved with this platform is enabling Mars MARS is our program to for modernizing and accelerating regulatory submissions.
[22:13] Regulatory submissions are the lifeblood for FDS review process. When a company submits an application for a new drug or device, FDA reviewers must analyze enormous volume of data, both structured
[22:25] and unstructured using clinical trials, results, safety data, labeling, and much more. The Mars is an AI powered initiative that use the data bricks platform integrated with
[22:37] Elsa as I said our own homegrown FDSAI platform to accelerate the review process. It's about getting the right information to the right reviewer faster, reducing the time reviewers spend on data wrangling and supporting
[22:51] better informed regulatory decisions. Halo and Elsa together represent what responsible AI can do with the human in the loop. AI does the leg work. AI augments the
[23:03] expertise of our scientists and reviewers but does not replace them. And here I just want to be honest with you all about what we learned along the way because I think this is where the
[23:16] real value is for anyone in this room considering a similar journey. The first lesson is a stakeholder engagement. It is not a one-time event. It is an ongoing commitment and every center has different needs.
[23:29] They have different timelines, different risk tolerances and continuous communication and alignment were essential. The second lesson was the security planning can never start too early in
[23:41] Fedramp high environment. Every architectural decision has implications. Starting security planning late creates bottlenecks that are very hard to unwind. The third lesson was build for
[23:53] flexibility not hard coding. We had integrations where specific settings were baked in endpoint URLs, version numbers etc. and when the platform we updated it
[24:05] cause issues. So building flexibility from the starting. The fourth one is wave wavebased migrations by center workload type bounded by risk and uh compounded
[24:20] learnings. The big bang cutovers don't survive like the 40 plus complex use cases with different workloads and risk tolerances. The last one is a strong vendor partnership. It is an operational asset.
[24:34] Data bricks was not just a vendor. They were partnered with us made made the the federal scale migration possible.
[24:46] So where we are headed, we are very proud of what we have built so far, but we are not done yet. We are expanding the serverless compute utilization across the platform and also we are leaning into AIBI and Genie spaces across FDA just like any other
[25:00] agency. We have so many reports and dashboards which we plan to reduce by end of year by at least like 10 to 20%.
[25:13] These capabilities will allow FDA employees to interact with data more intuitively without requiring deep technical expertise and we are growing the Elsa capabilities as well. We are expanding its capability to the user base across FDA by
[25:27] integrating more of FDA's data assets in Unity catalog. We are making Elsa smarter and more useful for reviewers and scientists alike.
[25:42] We also exploring the delta sharing to securely govern the data exchange with our partner agencies and also external stakeholders enabling more seamless collaboration while maintaining the FDA controls
[25:55] FDA controls and also we are expanding the model serving for machine learning workloads across the regulatory science programs. Our goal is to put the AI at the
[26:07] fingertips of our scientists securely and responsibly to support FDS mission. So the platform that we built is a is not a destination, it's a foundation for
[26:19] us. So I want to close with three things I hope you take away from this session. The first one is the authorization is an accelerant.
[26:31] Pedramp high sponsorship was significant inflection point for FDA for government organ organiz organizations in this room and those of you who work with the government authorization is an enabler
[26:43] not an obstacle invest in it proactively and it will open doors the second most important theme right now is the governance is the prerequisite for AI
[26:56] Unity catalog wasn't just our migration checkbox it made secure responsible scalable AI possible at FDA. You cannot build trust without governance.
[27:08] Finally, run in parallel. Don't wait for perfection. I didn't mention in the previous slides, but this approach actually enabled us to speed up our consolidation of data across the agency. It provided the perfect opportunity for
[27:21] us to onboard all our new centers and tenants while we parallelly started working to convert all our existing centers who've been using this environment into Unity catalog.
[27:33] FDA's modernization journey is a ongoing and I'm proud of what we have built with the partnership between us data bricks and deoid.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.