Timestamps

(00:00) Introduction to fileAI and Tim Prugar

(02:06) How Tim met founders Clare Leighton and Christian Schneider

(06:19) Synergies from being Head of both Product and Engineering

(08:52) What is Model Context Protocol? And what does it mean for the generative AI revolution?

(11:33) Use Cases for fileAI technology and future of Enterprise AI adoption

(17:43) The role of developer communities in driving Enterprise AI adoption

(19:29) Competitive landscape and moats for Enterprise AI platforms

(21:34) Remaining barriers to AI adoption for Enterprise

About our guest

Tim Prugar is the Head of Product and Engineering at fileAI, leading the technical teams who are solving the problem of AI data preparation.

Previously, Tim was the VP of Operations and Product Owner at Next Caller (YC14) where, as a member of the leadership team, he grew the company and guided it to a successful exit to Pindrop Security in 2021. While at Pindrop, Tim functioned as the Chief of Staff to the CTO, bringing scalable processes to Pindrop’s Product and Engineering efforts. Tim relocated to Singapore in 2022 to set up Pindrop’s Asia-Pacific and Japan office, functioning as the lead technical resource in the region.

Follow us on LinkedIn for more updates

Subscribe to our monthly newsletter for all the news and resources

Directed by Paulo Joquiño

Produced by Paulo Joquiño

The content of this podcast is for informational purposes only, should not be taken as legal, tax, or business advice or be used to evaluate any investment or security, and is not directed at any investors or potential investors in any ⁠⁠⁠⁠⁠⁠Insignia Ventures⁠⁠⁠⁠⁠⁠ fund. Any and all opinions shared in this episode are solely personal thoughts and reflections of the guest and the host.

Transcript

Paulo: Hi folks. Welcome back to On Call with Insignia, where we go on call with leaders innovating the future of Southeast Asia’s internet and digital economy and beyond, or as we like to call it, ASEANnovation. I’m your host, Paulo Joquiño, and we are back with a guest from a familiar portfolio company for those who have been tuning in.

We’ve had their founders, Clare and Christian on our show over the last two years (our first call with them, and then our second). Now, I’m really honored to have more of their management. I always like it when our portfolio companies come on the show, but they introduce to me everybody in their team and I get to have a more holistic picture of the company and hopefully you guys too as well.

We’re here with none other than Tim who leads engineering and product at fileAI. For those who are tuning in, I think in our last conversation with the founders of fileAI, it was still Bluesheets at the time. It’s been some time. There’s been a lot of changes.

They’ve rebranded into fileAI to better focus on their core functionality which is really taking unstructured data from various sources, different types of files and all that, and being able to create some kind of file intelligence that organizations can use to build out their AI workflows, have more automated workflows and a lot of other different use cases that we’ll talk about in this podcast.

With a lot of changes that have happened since we last talked with the founders, I think a lot of them center around the product itself. More recently after two or three years of delivering this product to different corporate multinationals from insurance companies to F&Bs.

Now they’ve made this technology a lot more accessible, a lot more public through their API and platform launch which was as of this recording a month ago. By the time you guys are listening to this, they would’ve launched a second layer of product on top of that, which is their MCP, which is like a cloud server.

But I guess I’ll let our guest describe that. I’m sure you can describe it a little bit better, but there is a lot of exciting stuff to catch up on. But before we get into it, thanks so much, Tim for coming on the show. Hope you’re feeling great getting into this.

Tim: My pleasure. Thank you for having me.

Tim’s Journey to fileAI

Paulo: Tim is currently calling from Singapore but he’s had a career that’s spanned quite a few countries as well. He’s only joined fileAI over the last year or so. Off the record, we were talking about how he had met Clare at Singapore FinTech Festival where incidentally I had been in Singapore at the time and I had visited the fileAI booth, but I didn’t manage to get to meet Tim at the time, but it was quite an interesting small world. Maybe we can start off our conversation with fileAI, like how you encountered fileAI, how you met Clare and Christian and decided to take the leap and then join the company.

Tim: It’s funny, we’re actually in the exact building where I met Clare. I met her at the Singapore FinTech Festival last year through an acquaintance and a friend of mine, and was really struck by the way she spoke about the business and the problem space, and was very impressed with her. By sheer coincidence, I happen to get introduced to Christian a few weeks later.

My background’s in startups and scale ups. This is my third AI and ML startup and I’ve had the privilege three times now of working for founders. Working directly for the people who took the risk and took the leap. I was really impressed with Christian and Clare, not only how they talked about their ambitious plans for the business, that art of the possible, that big audacious goal, but how they paired it with what I saw as a really responsible deployment of capital.

Those are the types of people that I want to work for who understand they’re not just trying to do something amazing and new and innovative. But also have to be stewards of the business and think about revenues and opex and all the things that go into being successful, and how strategy is a part of that as well.

I was looking for a new opportunity. I love to build. I like to build something from nothing and at fileAI get to build something from something. It was a great opportunity and I was lucky enough to be brought on board, so thrilled to be a part of it.

Building on Strong Foundations

Paulo: Speaking of building something from something maybe you can share a little bit about where fileAI was at the time that you met them, especially from a product standpoint, and were there already talks about doing this whole API launch and the rest of it.

Tim: When I joined there was a really strong customer base, a really strong revenue base, and a really strong replicable use case amongst the customers. The engineering team was in place, the go-to market muscle was in place and there was a really strong customer success and customer support function.

That kind of customer centric mindset, which is really sometimes hard to develop in a startup, was already in place. That was an exciting thing to walk into and why I enjoy working for and with Christian and Clare is there’s never complaints, complacency. It’s always what’s next? What can we do differently? How can we innovate? How can we be ahead of the curve?

There was this really incredible inflection point with the rise of generative AI and the way it was being brought to everyone in a way that never had before. To act on and innovate on that, to build the next generation of fileAI, which effectively allowed us to penetrate more use cases, more document types, more regions and bringing the customer mindset and customer obsession to that new platform.

One of the things we were hearing a lot from our customers was the wish to be able to self-serve more, to do more to, they were using ChatGPT, they were using AI services, they’re saying why do I have to go to someone to be enabled? Why can’t I do that myself?

The North Star of the V2 platform is flexibility and self-service and allowing our customers and providing them with the tools to build their own workflows, to build their own operations, and ultimately to find their own ways to use fileAI with our team as guidance and our team as the builders of those tools to really have a truly horizontal platform where large financial service institutions to insurance companies to logistics companies and everything in between can find a use case to build the excellence on fileAI.

The Product-Engineering Integration

Paulo: I think it’s a pretty meta development if you think about it. From you guys trying to automate workflows for your customers, you’re now trying to automate the product itself so that people can use it however they need it. What I find interesting about you coming into this role of fileAI is that you’re both VP for product and engineering, and oftentimes I think there’s a gap sometimes in a lot of startup teams, you have a team and then they introduce you like a product manager.

Then there’s sometimes it can lead to disagreements and timelines and workflows and all of that. How is being in charge of both those functions or do you really see them as like how you view the two essentially? And how has that helped with pushing out V2?

Tim: It’s a great question and I don’t necessarily have the right answer. I have my answer. I see them as very closely coupled. There’s definitely a school of thought that kind of product is the marketing revenue business side, and engineering is the technical side.

I just don’t align with that. I think about my career. I think engineers are excellent at listening to customer requirements, figuring out the best and most efficient way to do something. Unlock user experience and usage and consumption, which fundamentally is what you want from a product.

Being able to orchestrate the design function at fileAI with our forward thinking designers, our AI function which is our combination research and development cloud software engineering function, our product managers, our core software engineering functions, really working as one and having that constant communication and feedback loop.

I think gone are the days of marketing and engineering. This is what we think the industry wants: the speed at which people can iterate now with AI enabled tools and the velocity with which you’re expected to ship and iterate. Really having that tightly coordinated unit between product and engineering I really enjoy.

Then obviously bringing in the customer voice and the voice of the sales team. I think sales often get a bad rap in the AI space if they are the closest to the customer. They’re hearing objections every day. They’re learning new use cases, and we really have outstanding sales leadership and bringing those ideas to product to build, deploy, and iterate on. The engine of running both product and engineering while time consuming and not always easy. I think it is the right model for what we’re bringing to our customers.

Understanding Model Context Protocol (MCP)

Paulo: I think that the speed kind of speaks for itself. I think if my calculation is correct, you joined just after the Singapore FinTech Festival, which was in November last year, and as of roughly eight, nine months from then. You guys have managed to push out not just one product, but a second one, which is also the MCP. Before we get into it, maybe we can do a little corner, explain to me like I’m five what MCP is exactly. And how does that help?

Tim: MCP stands for model context protocol, and it was like most great engineering inventions come up with by a couple of frustrated engineers. Anthropic has this protocol and the idea is how do you get large language models to communicate with other plugins, API integration systems without necessitating these painstaking manual integrations?

If you think of a bookshelf is my favorite analogy. An API would be plugging into an individual book. You have to know what book it is. You have to plug right in. You’re limited to the knowledge in that book. What model context protocol allows is for you to plug into the bottom of the bookshelf and have an agent that now has access to all of the knowledge on the bookshelf so it can decide what is the right book. It will make mistakes. It might sometimes choose the wrong book, but it has that freedom and that versatility to learn from everything on that bookshelf.

At fileAI and Christian I’ll largely give the credit for this, of having the vision of this is something we need to run at. The dust hasn’t totally settled yet on how enterprises specifically are going to leverage agents of whether they’re going to use vendors to build for them, whether they’re gonna build internally, whether they want their agents reaching out or external agents reaching in.

That’s all kind of getting figured out right now, but our belief is that kind of at the core of this is going to be allowing that flexibility not only for agent builders, but also enterprise agents to be able to operate and interact with fileAI’s platform, however they see fit.

We’re coming at this from two angles. First is our developer community, who we know they’re building on top of MCP. They’re using Claude Desktop, they’re using Cursor, and we want to give them access easily to our full API suite so they can build their own workflows with fileAI embedded now at the same time on the enterprise side. We don’t want to wait and react to how enterprises are going to be using agents and agentic workflows. We want to be prepared so that we can drive that conversation. Then as we learn new information or new ways of doing things, we’re nimble enough to pivot and react to that.

But we’re really proud of the MCP server. It’s live now and we’re getting ready for a formal launch. I think it’s going to really strengthen not only our developer community, but also our ability to lead enterprise conversations regarding how to safely, securely, and reliably leverage agent workflows at scale.

Enterprise AI Best Practices

Paulo: You mentioned it’s not quite clear how enterprises are really going to use agents at scale. Then you mentioned the trifecta of safely, securely and reliably. Maybe you can share a little bit about what you are seeing enterprises. How are they? What are some of the best practices or use cases that you’ve seen so far from fileAI customers and how they’ve used fileAI technology?

Tim: I think we saw when ChatGPT really exploded onto the scene, a massive number of enterprise first movers who wanted to, either through customer facing chat bots or things started building these agentic flows and leveraging large language models and generative AI.

There were some missteps early on and people pulled back a little bit on that front of what does this look like customer facing? What we’re seeing within our customer base, one is largely for internal processes where there’s a really strong ability to reliably test and iterate and learn over and over again.

The second is granular control of privacy and access. Making sure the agent is only touching things it has been definitively given the privilege to not allow to run wild. I think the third is setting really strong success and evaluation criteria. Rather than seeing what happens and hoping for the best, thinking like what does excellent look like, feel like, and sound like as we’re running this test and doing this.

Then I think the fourth is identifying areas where data has been sitting that’s been untouched. Some of our customers want to use AI to automate existing workflows and that’s perfectly fine and requires quite a bit of documentation of the flow. Figuring out the tribal knowledge, what’s written in Confluence versus what lives with some guy named Dave in the basement.

Then the other end is customers who are looking to completely reimagine their workflows. We did it this way for 10, 15 years, but we want to see how AI can make us think about this completely differently. I think that’s the most exciting place to be particularly with agentic flows. But those are just some of the best practices we’ve learned so far.

At fileAI we’re working with agents too, and we’re experimenting and learning and seeing how we can improve our processes. Part of that too, I think, is putting yourself in the shoes of a customer, if you’re going to have successes and you’re going to have spectacular failures, and thinking about what that feels like and how you can learn from it to really build a bulletproof system.

Tim’s Favorite Use Case: Multi-Document Validation

Paulo: As a product and engineering guy yourself what’s been your favorite kind of use case for an agentic workflow?

Tim: My favorite use case has been validating a number of seemingly disparate documents against a set of rules. That sounds a little bit unexciting but let me explain the applications. If I think about an insurance claim, for example, I might have an x-ray, I might have a medical certificate, I might have the claim form. I might have a handwritten note from my doctor. I might have photographs of where I fell or had an accident. There’s a host of things there.

Being able to leverage agents not only to classify each of those documents in one shot, so knowing what they are despite the system never having seen them before. Being able to extract information from images and text, whether it’s handwriting, whether it’s a western language or not, whatever the case might be.

Being able to synthesize that information and then compare it against a prompt that says approve or don’t approve based on this criteria. Being able to leverage agents not only to figure out what APIs to call, how to classify that information, where to store it, how to read it, and then ultimately how to synthesize and make a decision with the biggest thing being explaining reasoning and rationale.

This was a non-negotiable for Christian and Clare was the idea of building trust and confidence in the information that fileAI returns. It’s not enough to say, here you go. Especially in an enterprise context. People want to understand where this information came from. How can I be sure it’s the information I gave you and how can you defend your decisions here? What is the rationale so that I can go back and check if need be.

Being able to layer that onto the work that we’re doing with agents and we’re seeing our customers do is a really huge value add. It’s effectively updating machine learning confidence scores for the generative AI revolution.

Paulo: I think that’s super exciting. As somebody who uses agents a lot I think that’s something missing from the consumer side. You get a lot of this information, but you don’t really know if it’s real or not, or where it’s coming from, and you have to go through these iterations to make sure you’re not being gaslighted by your ChatGPT or something like that.

Tim: A hundred percent. I think you mentioned the Singapore FinTech Festival earlier. I think another really neat use case that actually one of our sales leaders leverages is business cards from a conference. Uploading those into fileAI, having them automatically categorize, extract, and then being able to build an agentic flow that helps you figure out if you’re traveling, who in that Rolodex is in a particular city.

What did you talk about? If you have a note section and pull that all together into a recommendation and build actually a trip plan for your business trip of who you can see where, when. Pulling from location, even pulling in weather information, public holidays, all those kinds of things. Using that agent as a revenue driving assistant is a kind of neat, fun use case I’ve seen as well.

Paulo: That would be super cool to have as a tool on your phone. Like maybe something that’s even like location detection related. As soon as you land, it’s oh, here are some recommendations for who to meet.

Tim: I think that’s the main driver behind the MCP server too, is that if you can think it, you can build it.

Developer Community and Product Development

Paulo: Speaking of building, I wanted to touch a little bit on the whole, you guys have launched V2 on Product Hunt and are also gonna be launching, I guess as of the release of this episode. Not quite yet, but you guys will be also launching the MCP on Product Hunt as well. You’ve talked about how important developers are to the product itself. Maybe you can relate how developer activity using fileAI has actually helped you in terms of the roadmap as well and also maybe even making more headway into other enterprises or other kinds of industries.

Tim: Absolutely. I think two main ways that the developer community has really helped fileAI first is just direct, consistent, no-nonsense product feedback. If something isn’t optimized or they would do it a different way or anything we hear about that and it, one you can’t hide anything in that context. But two it allows us to get really actionable feedback to put in a sprint and iterate on right away. That, I think, is a big thing.

The other is it really helps us be aware of the landscape. When you’re hearing, oh I found a lot of success plugging into this particular platform, or I use this tool. That might be something we weren’t aware of. We start thinking of new opportunities for either partnerships, integrations, revenue, and really spot some up and coming ERPs or legal document processing systems that we might not be aware of. Keeping our fingers to the pulse in that way helps drive our innovation.

It helps us keep ahead of the curve and frankly it helps us from being complacent, which in startups is the absolute killer, is the second you take a deep breath and say, okay, that’s good enough. That’s when someone tends to be passing you by.

Paulo: I think it’s really interesting how you’ve built it like this in two ways, for developers as well as for enterprise. Because obviously I guess if you’re only just focused on revenue and all that, then obviously some people might say, why don’t you just focus on enterprise? But then the developers actually put you guys ahead of the pack when it comes to thinking about what to build next and where the technology’s going.

Competitive Advantages and Market Position

Paulo: Speaking of keeping competitive, I also wanted to touch on, obviously it’s in your tagline, the world’s largest horizontal file intelligence platform and all of that. The technology, obviously, competitors are just a click away, so to speak. There’s a lot of products emerging, maybe not exactly horizontal as you guys, but maybe within a specific industry. They’re building for legal or they’re building for healthcare or something like that. How do you view the competitive advantage developing from a product perspective?

Tim: Absolutely. fileAI has a significant moat in the number of years we have already spent processing what I call in real life documents at scale. We have cut our teeth doing the handwritten medical forms, the blurry photographs of gas bills and traffic tickets. We’ve moved beyond what is effectively the, I want this research paper and marked down use case.

Having that expertise having built models specifically for those kinds of use cases and spending the bulk of our early stages as a company working in some of the more difficult languages and the more overlooked languages as these machine learning and AI advancements were coming out.

There are a lot of countries in our region that were simply left out or left behind, and having the wherewithal to recognize that, to prioritize that Thailand is one of our target markets. We’re expanding globally with an office in New York City and servicing US businesses that have a heap of documents that they need processed and digitized and extracted from an unstructured to a structured format.

I think having this experience, understanding the processing firepower and the model firepower you need to meet some of these use cases that actually exist in the wild. Really being able to service languages across regions from not only a performance perspective, but also a regulatory and compliance perspective is a massive advantage for us.

Overcoming Adoption Barriers

Paulo: I guess like from how we’ve been talking about fileAI, I remember there was a phrase that Claire used to say oh, it sounds too good to be true. But that still remains the case. I’m curious, like what, in your view, remains the biggest barrier to adoption? For a platform like fileAI? What are, what have you heard from the market and how are you guys planning to resolve that or overcome that?

Tim: I think there are two significant barriers, one of which is not fileAI specific and one is. In general, customers’ awareness of fluency in and expectations of generative AI and large language models just varies so much and there’s a massive education piece. Both kinds of understanding where the customer is in their journey and how to position, deploy, develop our services in a way that meets that need. There’s just a huge chasm in what folks expect and need from technologies like fileAI or a host of other AI firms at this juncture.

I think you’d probably, if you had 50 people sit in this seat you’d probably hear at least 49 say the same thing. I think the second is the need for high accuracy. The most successful AI firms at this juncture, in my opinion, their products are largely like that’s close to what I need. That code is good enough. That product prototype is good enough that shows what I need. fileAI’s customers have a justifiably high bar for excellence and accuracy, and we endeavor to meet that. We don’t get in arguments with customers about whether 80% accuracy is good enough. We talk about our plan to get to 100% and deliver on that.

I think that is the barrier, the largest barrier to adoption is making sure we’re aligned on and meeting that high accuracy threshold and we get over it for our customers. We meet those accuracy thresholds we have, our CS function sharing frequently, our model benchmarking and accuracy on the actual customer’s data to build that confidence that we are always moving in that direction.

But I say that the toughest hurdle is really making our product, our models, our conversations all aligned in this bar of what we expect of ourselves, which is at or near perfection on our extraction and form filling.

Paulo: You don’t know what you don’t know. I think it is great that the customers are letting you guys know. I remember a conversation with another product like CCPO and she said the best customer to have is the one who complains. It is always good to have feedback.

Tim: The feedback is invaluable for continuous improvement and building trust with our enterprise customers.

Stay tuned for part 2! Follow us on LinkedIn or subscribe to our YouTube to get notified!