> ## Content Index > Fetch the complete content index at: https://blog.ai.gov.sg/llms.txt > Use this file to discover other available public pages before exploring further. # One Year in Multimodal AI at GovTech - URL: https://blog.ai.gov.sg/one-year-in-multimodal-ai-at-govtech/ - Published: 2026-08-07T00:00:27.000Z - Updated: 2026-08-07T00:00:26.000Z - Description: “I came here expecting to build models for people who already knew what they wanted. I discovered that the harder work is figuring out what should be built, who it is for, and whether it solves a real problem.” - Author: Sumiko Teng - Tags: The People, Multimodal I'm Sumiko, a data scientist at AI Practice. I joined fresh out of school in July 2025, so this post lands almost exactly one year in as a tiny reflection of my time so far. When I first started, I thought I knew what the job would be. Training models, annotating datasets, running experiments to push an accuracy number up. I assumed the problems would be fairly traditional multimodal/computer vision ones, which felt like a natural continuation of what I'd done in school. I was genuinely excited about that, because I'd enjoyed the work a lot. I was satisfied with what I delivered in school and I just assumed more of the same was coming. What I didn't expect was how quickly building the model became only one part of the work. In school, the scripts and the models I built had a lot of potential impact, and potential was the keyword. GovTech put me on the ground with real problems and the real people who have them, and that turned out to be a different thing entirely. Solving those problems isn't only about the models and the data. It's about understanding and empathising with the end users, why this is a problem they need solved, and how AI can help them deliver better. The field has also moved incredibly fast since I was in school. Things that would have taken an entire project back then are an API call away now. Part of the job is staying afloat of what has only just become possible, working out why it's possible now, and looking for the places where a new capability meets an actual problem. **That combination, working with real problems while the technology itself is changing rapidly, is what has shaped much of my role.** ## **Greenfield exploration and prospection** One big part of my work is greenfield work. A new technology comes out, we want to validate it, and we want to find out whether the capability could genuinely uplift the way people across government work. That's harder than it sounds, because it sometimes means exploring a technology before there is an identified concrete use case. I explored AI-generated presentations and videos, using Paper2Video, an open source research pipeline generating LATEX slides and speaking avatars. I wanted to see how far I could take it, so I gave it a photo of myself and a recording of my voice, generating a video of me presenting. It was tricky. There were uncanny valley issues, and it also wasn't clear how anyone across government would actually use it. We showed it around, and the challenge turned out to be finding someone who genuinely needed it. At some point the honest answer was that this probably wasn't good or ready enough to solve anyone's problem, so we moved on (partially). ## **Prism** We did not give up on the generative media space and shifted more attention towards infographics and presentations. That exploration became Prism. My buddy and I decided to take the generative media idea to more users through the [Build Hackathon](https://build.tech.gov.sg/?ref=blog.ai.gov.sg). ![](https://storage.ghost.io/c/ca/f5/caf570e4-5b80-4fc2-b4c8-0c544c2ce19f/content/images/2026/08/data-src-image-80729524-1266-46c5-a123-6dda3eb8c2d3.png) Group photo with the Finance team who kindly “dogfood” Prism! Making slides was a problem a lot of officers had, making and that was also what made it hard to scope. Every officer puts together decks and infographics at some point, which means there's no single group to optimise for. The hackathon turned out to be a good way in which we got to talk to different profiles of users. The validation was apparent when we won the judges' prize and best pitch. ![](https://storage.ghost.io/c/ca/f5/caf570e4-5b80-4fc2-b4c8-0c544c2ce19f/content/images/2026/08/data-src-image-fc2d3da8-d1a0-405d-b2c0-1e051942ba14.png) However, more useful than the prize was putting Prism in front of users across agencies via GovTech, allowing us to see whether it survived contact with their actual day to day work. Building Prism was the cheap part. The user research, the feedback and the improvements we made afterwards were what shaped it into something that solves a real need. ## **Grounding exploration in real problems** A large part of my time here has also been spent working directly with agencies. The work is not always purely technical. We spend a lot of time clarifying what officers actually need, separating assumptions from what we know, and making sure the solution stays pointed at the problem rather than at the technology. ![](https://storage.ghost.io/c/ca/f5/caf570e4-5b80-4fc2-b4c8-0c544c2ce19f/content/images/2026/08/data-src-image-fcbfcb66-4b83-48a0-ab08-9e534d768d0d.png) Joint project with NHB presented in CAA 2026 I've found that I enjoy talking to stakeholders and digging into their pain points, and it lets me deliver something useful for them. It's one of the things that makes this job feel meaningful, and it matters far more than I assumed it would. ![](https://storage.ghost.io/c/ca/f5/caf570e4-5b80-4fc2-b4c8-0c544c2ce19f/content/images/2026/08/data-src-image-add193bf-c2fc-4dd9-8e31-c0f19aa8a7e6.png) Launching the multimodal AI handbook and Platform AI Studio at the AI Champions Bootcamp Out of those agency use cases, I also build prototypes despite not having formal full stack experience. Many of the officers we work with are domain experts rather than developers, so a script or an API alone is not enough for them to evaluate whether a capability is useful. If we want them to find out whether an AI capability can solve their problem, we have to make it easy for them to see it. That's why we brought [PlatformAI Studio](https://studio.platform.ai.tech.gov.sg/?ref=blog.ai.gov.sg) into the picture. It is a playground and sandbox to try out AI capabilities and it was great to see how many people were interested in testing multimodal AI on their own data. Engaging agencies then surfaced a different gap. Officers wanted to try multimodal AI, but unfamiliarity with the space made it hard for them to know where to start. That's where the [Multimodal AI Handbook](http://go.gov.sg/mm-handbook?ref=blog.ai.gov.sg) came in. It's AI Practice's amalgamation of *what has been done* and *what actually works* for multimodal AI in the public sector, so that officers can find their own way rather than us sharing to different teams in isolation. ## **Sharings and the community** ![](https://storage.ghost.io/c/ca/f5/caf570e4-5b80-4fc2-b4c8-0c544c2ce19f/content/images/2026/08/data-src-image-16a2fdbd-14cb-4ca2-9d23-f5fe60527a74.png) Exploration also rarely happens alone. One of the more valuable parts of this year has been the AI and ML sharings we run fortnightly. Different officers present whatever they've been working on, whether it’s a work project, a side quest, a hobby, or just prospection of some new technology. It pulls in people from the Forward Deployed Teams and other parts of GovTech, so we find out what everyone else is doing and learn from each other. Prism was demoed twice while it was still a work in progress, and exactly what the sessions are for. A low stakes platform to put unfinished work in front of people, hear honest feedback, and get ideas for the next step. The socials at the end of these sessions have been just as valuable for getting to know people and keeping the workplace fun. ![](https://storage.ghost.io/c/ca/f5/caf570e4-5b80-4fc2-b4c8-0c544c2ce19f/content/images/2026/08/data-src-image-40699a93-cc14-4925-8580-d5c4014d0665.png) ![](https://storage.ghost.io/c/ca/f5/caf570e4-5b80-4fc2-b4c8-0c544c2ce19f/content/images/2026/08/data-src-image-e86a2040-c184-4759-bb34-3674f1e8be7a.png) ## **One year in** Looking back, the job has been much broader than building models and solving data science problems. It is about spotting where a new capability might matter, testing it before there is a clear playbook, and translating what we learn into something people can actually use. That also means being comfortable with uncertainty. Not every exploration becomes a product, but even an idea that does not work can narrow the search, expose a limitation or point us towards a better direction. A year in, and that is the part I find most exciting. The technology keeps evolving, and every meaningful improvement opens up problems that could not be addressed in the same way before. My job is to explore those possibilities early, work out which ones are worth pursuing, and help create a path from emerging technology to tech for public good.