Transformative Al & Technology for Decisionmaking Part 3

Welcome to “HSDF THE PODCAST,” a collection of policy discussions on government technology and homeland security brought to you by the Homeland Security and Defense Forum

In this episode, we lay out a practical path to trusted AI in government: build a mission fabric that captures operational context, invest in shared data architecture, and measure decision quality through iterative delivery. We also tackle brittle models, multimodal sensing, and smarter procurement that values outcomes over promises.

Featuring:
- CDR Jonathan White, Cloud and Data Branch Chief, U.S. Coast Guard 
- Carin Quiroga, Chief Data Officer, Immigration and Customs Enforcement
- Courtney Whelan-Stillmun, Principal Architect, Google Public Sector (moderator)

This discussion took place January 22nd, 2026, at HSDF’s Technology Innovation in Government Symposium

Follow  HSDF THE PODCAST and never miss latest insider talk on government technology, innovation, and security.  Visit the HSDF YouTube channel to view hours of insightful policy discussion.  For more information about the Homeland Security & Defense Forum (HSDF), visit hsdf.org.

Transcript

Defining AI Success In Government

SPEAKER_00 0:00

So so looking three to five years ahead, what would success look like for AI-enabled decision making in the government?

Building Trust And Mission Fabric

SPEAKER_04 0:08

So for me, success is speaking of that those two lanes, right? Going from explicit AI to implicit AI. Right? And and trust is uh is the normative factor on that. Uh and and so I think in the next three years we really need to focus on building that trust factor, right? And when you're when you're seeing it the um the AI agent doing its thing, you should the thought that comes to your mind should be it's gonna produce something awesome. Can't wait for that to happen, right? Not, uh, I don't want to review that thing, it's probably gonna take a couple cycles, right? So I think we really need to we need to build that trust factor, and and a lot of that is is trial and error, a lot of that is is quality data, right? Because the AI is only as good as the data that you give it. And then I think the other the other thing I want to see in three years is this uh complete integration of the mission fabric, right? So something on top of just the data fabric, but there's there's all this operational data. And a lot of that data gets thrown on the floor, right? We don't we don't end up using it or it gets turned into like slick notes or some sort of free text uh stuff. So how can we capture all of that information and really derive a lot of insights and in real-time uh synthesis of that information? You know, there's there's a future where um you know we're viewing the Coast Guard battle space, if you will, right? All the cutters and aircraft that are out there in the world, and what are they doing at any given moment, right? I don't think any you know nobody can really answer that question very well. Like you have to be on the cutter to really understand, or you have to be on the aircraft. So if you can say, like, what is the Coast Guard doing today at any given moment, and you'd be able to answer that with fidelity, that would be really awesome, right? That's a true common operating picture, a true operational space management. So I think in three to five years, as you start collecting all of those operational threads, you have great connectivity, you have storage, you have processing, you have all these things lining up with each other. You can start doing things like that at the aggregate and scale level, which changes how you do operations, right? That's fundamental change, that's transformative change in how you do operations. So I'm really excited about that journey.

Coast Guard’s Common Operating Picture

SPEAKER_01 2:12

And for me, for three to five, I'm also gonna focus on the data, like having that. We don't have this data architecture yet in place. We've it hasn't been funded. I'm a team of two data people. Um, so for success for me is really in three years is having that infrastructure and architecture, not only for us, but for the whole entire agency and also federal government. Because as you were mentioning that mission data today, how it's kind of put on the side, we're actually now at the point of needing mission data from all of the components and agencies to be able to do our job, not only them, but state and local and everybody else. And so, you know, in three to five years, success would look like that we have the platform, we have the technology, we have the data products, we have the AI, and everything's in place for us to be able to function interoperably with each other seamlessly. We have processes in place to share and document that is very easy and streamlined, doesn't take six months to approve, and things in place. That's where for me, success would look like for me. And the data with bodega would be like full of all kinds of data from full stuff.

Agency-Wide Data Architecture Goals

SPEAKER_00 3:14

Full shelves. Yeah, I mean, that's what we're we're seeing in the industry as well, is a big push toward, you know, multimodal analysis, looking at video feeds, looking at uh situational awareness around a particular satellite imagery or what have you, moving into that multimodal component, um, as well, of course, as the push for agentic AI and being more proactive. And that's really what we're seeing a lot of interest in as well, with some of the people the customers that we work with around, hey, I want to move from you know my user dictating what's happening to your point to more the agent going out and doing things for me, and then having me kind of sign off and hit that checkbox, yes, this looks good, type of uh action.

SPEAKER_01 3:56

So and we would have the focus on where they need to be, right? The human and focusing on those areas in AI, doing the other areas that they don't need to, and really having building that trust. So trusting it and having AI as an ally. And I think in three years, I know we can get there with confidence in AI. I'm confident we'll get there.

SPEAKER_00 4:13

I love it. Even me.

SPEAKER_04 4:16

I'm with you, yeah.

SPEAKER_00 4:17

We'll be there. We will be there. Um so I think we're able to take a few questions from the audience, and yeah, so I think there's some some mics coming around, but got their hand up in the middle here.

SPEAKER_02 4:38

Can you hear me? Oh, yeah. Uh Commander White, this one's for you. Um, thank you so much for sharing. I really appreciated your uh focus on governance and data sharing. And I'm curious um if the Coast Guard is thinking about a future where models trained on Coast Guard sensor data could be reused or adapted by joint partners as part of a broader reciprocal maritime domain awareness ecosystem.

Multimodal Sensing And Agentic AI

Audience Q&A: Model Reuse Across Partners

SPEAKER_04 5:04

Yeah, that's a great question. And and I think it's uh it goes both ways too, right? So the Coast Guard absolutely wants to use joint partner technology, right? You know, we're we're a heavy use user of Navy technology. Um I've done that through my career. Um so how do you integrate uh something that may have been built for a different purpose into the Coast Guard mission set? Um so it the problem with training models is sometimes they get really brittle, right? So they're very sensitive to the where the sensor might be located, what the context of that is, the sensor uh fidelity itself. Um and so I think sharing sharing that might be a little challenging because you you don't want to take a model and then put it in the wrong context and then it's no longer valid. Um and I think the way we train AI right now lends itself towards being pretty brittle in the trained model sense. Uh so uh I'd want to see a little more action in the research realm to make AI a little less brittle in that sense. I think generative AI makes it a little bit better, especially for object detection and scene description and things like that. Um, but uh you know, true, like, hey, put a box around this thing and and then you know, figure out what it's doing and all that stuff. Like I like to see that as a delivered package, right? So when you go out and buy a commercial off-the-shelf product, which may be boutique to the particular defense or or or um security mission that you're in, it should work. It should just work, right? It they should have already thought about all those contexts. So I think um for me it's more about validation of that. So if another mission partner has deployed these sensors and they're working really well, we should be buying more of those things, right? And we should be diverse, we should also be diversifying the the spectrum with which we are sensing as well, right? It's not just visual, right? There's radio spectrum, there's space, there's um uh indirect sensors, maybe data feeds coming in from commercial entities uh that are shared that we can buy that data from. So I think there's a there's a broader answer there beyond just maybe just boutique training of models. But I agree with you, we should be sharing uh our at least our experiences, but also if we can share the models themselves and be able to deploy them uh in context. That'd be great.

SPEAKER_00 7:25

Any other questions? I got one in the back here.

SPEAKER_05 7:41

Thank you all so much. Um you both mentioned um that you know some of the challenges around facilitating gridding uh data sharing and um you know enabling both the technical sharing across organizational and technical boundaries, um, but also ensuring that you had awareness of data quality across those different systems. Um could you break down some of those challenges a little bit more? Um, you know, are there challenges with ensuring that procured systems have the right uh capabilities? Is it ensuring internal engineering capability to do those uh things, or you know, are there other bureaucratic hurdles that are preventing some of that? Thank you.

Data Quality, Standards, And Sharing Hurdles

SPEAKER_01 8:24

I think it's all of the above. Um from what I've seen, and a lot of it is again, like I mentioned, peeling the onion to identify why those problems exist. I mean, we've seen things where development teams allowed a freeform field for state. You know, we have 50 of the maybe 51. So like it we we saw Virginia in the database spelled wrong. Like, so those are examples, and we know the tools and the products have that capability. Um, but then there's also been things like with new new tools and not having, you know, them having the ability to do certain things that we want to do or customize it for things that we need. Sometimes we do need to think out of the box, right? And then those cause tooling issues. So I think for us it's all of the above. Um, I think for us trying to get ahead of that is to get like the data quality tools to be helping us to monitor that, put those places, put the standards, put the requirements in place, and then like try to help, you know, monitor that so we can be proactive to make it better. And then normalize it right before we do the data products, not necessarily do we have to fix it everywhere? Like it's gonna be really hard on some of the legacy systems to make that improvement, but at least if we understand what it is when it's coming into the place when we're setting up the data products that they could be have better quality for use.

SPEAKER_04 9:34

Yeah, I think the um, you know, I I look at it from the technical side mainly is just just getting the data to the right place with security is is sometimes difficult, right? We operate on DOW networks and not DHS networks. So they're in there in a in itself is a is a barrier, right? Now we have means to to do that, but it's just extra steps in order to move the data from where where it resides and where it's controlled to a to a different control point. Uh and and I think we know we went down this journey uh recently where we the components said, Hey, we want to we want this data set. We really want this data set, like just give it to us. And I the first question that came out of my mouth was, what do you want to do with it? And it wasn't because I didn't want to give it to them, right? It was like, I want you to understand like what what is the purpose of giving you the data. And I think what we need to really get good at is answering those questions and so that we can give you the right information with the right, even the right standards, like meters, feet, centimeters, whatever. You know, all these all these little things that people are like, oh, don't worry, it'll be fine. It's no, but when you actually put it in the actual model or you you actually put it in the app, it's not fine, because if you didn't do the transform, you're you know, the quintessential example is the was the Mar the Mars uh lander that just crashed right into Mars because they went from meters to they didn't do meters to feet, right? Wow or meters to yards, right? So it was off by just a smidge, but that smidge caused the loss of the aircraft. So I think we need to be very, very careful about uh when we say sharing, it's way more than just a an agreement, right? It's absolutely all the way down to the schema and the data itself.

Prioritizing Multimodal Use And Talent

SPEAKER_03 11:20

Yep. Hi, thank you so much. I guess kind of to that point when it comes to looking at data sharing, um when it comes to and circling back to kind of the multimodal types of context building and situational awareness building, um, that happens in kind of different parts of DHS and just broadly understanding on the NATSAC front. Um where do you see kind of the priority being in that multimodal using multimodal types of data? Um is it like building those kinds of tools in-house? So looking at geospatial awareness and building talent within the government, or is it looking more at like lateral sharing? Are there priorities in that domain to being to bring in more multimodal types of data?

SPEAKER_04 12:14

I'm not really into that space particularly.

SPEAKER_01 12:18

Yeah, I mean, I would say, I mean, all the again, I don't want to say all the above for everything. I think that there's definitely a use of getting data and also the expertise from industry, right? And out there, I there's internally there's things like you mentioned that you're gonna have the expertise that others won't because of the nature of the job, right? And so I think that comes to play, but I think also just looking at you know getting it from each if if the data is available one way or the other, they're gonna try to figure out every method to get it if it's needed, right? And so I think it's just a mix of those, all of them brought together. But I mean to to your point, I'm not as familiar, especially on the geospatial side. Yeah. I know. You're more water and air forces, you know.

SPEAKER_04 13:02

Yeah, I kind of just go back to you know, what are you what are you using the information for? Are you extracting insight? So if you have an image, what what insight are you extracting for that image? Right, besides looking at it. Are you looking for particular objects? Are you looking for scenery? Are you looking for contraband? Whatever the case may be, right? And so depending on what that answer to that question is, you're gonna have to do a different model for that, right? You'll have to ask different questions. So I think that it goes back to like what are you gonna do with this information? And then the format kind of blends into the background, because that's really uh that's really an interface standard, right? There's an image or audio or video, it's really just interfacing. And what I'm actually gonna extract from the signal from that image is is is the important piece.

SPEAKER_00 13:47

And I think we probably have time for one more uh over here.

SPEAKER_06 13:52

Thanks very much. I'm Jeff Harnold, CIO at ERT, unearth science and space technology data and modeling company in Tyson's. Thanks very much for a terrific, really engaging and informative panel.

unknown 14:06

Oh.

SPEAKER_06 14:08

The altitude help. Um Thanks very much for our great panel. And um I was particularly excited because we got not only uh descriptions of use cases, but of actual running implementations. And Corinne said um something super interesting at the beginning, and that was that she's got proof by looking at better decisions that come out of the end of it. And given what we heard in the procurement panel this morning, the about the move from static statements of work RFPs to actual outcome-determined descriptors for contracts. Can you say some more, either of you, both of you, say some more about how you measure that that improvement in the decisions themselves? It can be really tricky to do it. Do you do it with some data withholding experiments? Do you do A-V testing across the decision itself? Can you actually do it in a way that you could put up with a bad decision if it came out, kind of a thing? Because I can think of lots of problems of trying to look at how you measure whether the decision itself was improved. Thanks.

SPEAKER_01 15:12

That's very interesting. Because I've never thought of it in that way before, as because I unfortunately wasn't here this morning, but from a contracting perspective, but it would be nice to see that shift of like demonstrated results. And of course, it all leads to mission value, right? With everything we're doing with AI, how is it impacting the mission? Are we reducing time? Are we doing more improvements? Are we doing faster decision making? So for me, I mean, listening to it from procurement and how that would happen and that demonstrated would be hard, like hard for me to think of on the fly, but I definitely think that's a you know a great concept to look at and think about.

SPEAKER_04 15:44

I think the I think the interest what I was thinking about when you asked this question is do we do experimentation in the government? Um typically we go through a planning process and we say, that's it, deploy it, right? Make it so, and and we kind of accept whatever comes out of the other side of that, which probably speaks to the static RFP problem, right? And uh maybe we need to shift left a bit on our our thinking of even on business process and decision making, right? How can we A B test this or blue-green test these these options? Do we have enough um do we have enough uh scale in order to do that testing, right? Do we do we can do we meet the P threshold, right? Uh the NP threshold. Uh and and that's that's usually a challenge in government because our scale is very, very tiny in comparison to worldwide type population dynamics, right, and doing those things. So I I'm I'm kind of torn on the answer, right? I'm like, hey, maybe we just just step on the gas and go and then know when to bail or know when to sh when to pivot, right? Instead of maybe doing a more academic journey. Maybe we need to build some more maturity in the data plane first before we can do the academic rigor necessary to truly understand what is better, what is worse, right?

SPEAKER_01 16:58

And I also think like just going back to the contract, I know we've done before in the past, like contracting where there's a like a phase one is a proof of concept or pilot. And it's like, hey, if that is successful, then we're gonna move on. And if you can't show that measurable impact and to the mission, then it's a cut deal, we close it shop, we go the next. So I'm sure there's ways, because I know we've done contracting before where it's like the phase one is like a pilot or a prototype. So I'd say that would also be.

Pilots, DevSecOps, And Iteration

SPEAKER_00 17:23

I think it does start with, to your point, Corinne, like the mission objective. Like what is the actual problem that we're trying to solve? Because we're not just trying to like do stuff with AI for its own sake. And then it gets into okay, well, what are the approaches to solve the problem and which of those approaches is the best? And that's really where we're talking about our POC phase, our iteration. And then once you determine, okay, based on all these tests, based on our scope and our tri success criteria, what is the best option? That's where we start thinking about, okay, now how do I incorporate that into an application if it if it's going into an existing application, or how do I deploy a new application? And to me, at that point is where we start going into the dev sec ops process and start thinking about canary, blue-green, AB testing and kind of implementing that to a base of users to see if it actually does work in production or if you do need to make a shift if that POC testing wasn't enough, and then maybe you do come back to that point and continue to iterate. So that that's how I've experienced it and looked at it from the perspective of those that I've worked with, but would love to understand if that makes sense to you.

SPEAKER_01 18:25

Yeah, and I know we only have a minute left, but like I also think it's it's for AI, it might be to do something like that later down the line, because I think we're still learning and we're still prototyping, and we wouldn't want to punish a vendor or an industry from like, hey, it didn't work because you know you we didn't have it on place on the government. And so, you know, that specific use case, we had a team that their goal was AI, and so that was just the function, the scope. So we were, to your point, it was failure. It did like there was tweaks we had to do that, it was failure up front. We had to modify and change it to get to success. And I think as we're growing and in this you know new place, I think that's the right approach. And I think in the future, after we get the confidence, we get our governance for AI, we get everything in place, moving to more of a model of like, hey, demonstrate this success would would make sense.

SPEAKER_07 19:10

I promise bring it into pipeline, it's because I'll be able to come out on the data side. But um, my problem we're gonna say was that the decisions are going to be happening.

Confidence, Governance, And Faster Decisions

SPEAKER_01 19:22

So yeah, they are. I got what you're saying. So yeah, and so we've we've seen how much faster we've been able to make decisions and how they have more insight because AI is highlighting the anomalies and the differences. But we now I don't know if, and I'll have to check on this, if they went back and like re ran some of those decisions in the past they made to see if like decision making got better. But that's a good point. So maybe I will bring that back and test them on it. But thank you.

SPEAKER_00 19:57

All right. Well, thank you, everyone. Really appreciate the time. I think we are at our time for today.

unknown 20:03

Thank you.

SPEAKER_00 20:03

Thank you all.