Portrait Analytics: From GPT-3 Experiments to $10M in Funding—The Pre-ChatGPT AI Journey

Building an AI company before AI was cool required either incredible foresight or foolish optimism. Sometimes both.

In a recent episode of Category Visionaries, David Plon, CEO and Co-Founder of Portrait Analytics, an investment research platform that’s raised $10 million in funding, shared what it was like to fine-tune GPT-3 for institutional investors in 2021—when most people had never heard of language models and the idea of AI as a thought partner seemed absurd.

The Stanford Basement Phase

David wasn’t looking to start an AI company when he enrolled at Stanford Business School between 2015 and 2017. He’d worked at several investment firms and loved investing. But something else was happening at Stanford that would redirect everything.

“Deep learning is kind of getting into the Zeitgeist and the Transformer paper which underlies a lot of the foundation behind language models today had just come out,” David recalls. The Transformer paper—the 2017 research that would eventually power GPT, Claude, and every major language model—was just academic curiosity to most people.

But David saw something specific. “I was just very curious about the idea of large scale pattern matching that machines were doing being similar to the large scale pattern matching that great human investors do.”

This wasn’t a random connection. David had spent his career watching how exceptional investors processed information—synthesizing vast amounts of data, spotting patterns across companies and sectors, building conviction through structured analysis. The parallel to what these early models could do was impossible to ignore.

Teaching Himself to Build

Here’s where most non-technical founders stop. They have an idea, they understand the problem, but they can’t actually build anything. David took a different path.

“I’m certainly not an engineer by background, but was just very curious,” he explains. He spent time at Stanford “taking independent study and teaching myself enough to be dangerous with respect to using machine learning models.”

The phrase “enough to be dangerous” matters. David wasn’t trying to become a world-class ML engineer. He was trying to validate his intuition through hands-on experimentation. This created something rare: a founder who understood both the domain (investment research) and the technology (language models) well enough to see what others couldn’t.

The GPT-3 Validation Moment

For years, David kept poking around with models. Then GPT-3 launched its API, and everything accelerated. This was 2020—still two years before ChatGPT would make AI mainstream.

David started fine-tuning GPT-3 for a specific use case: earnings call analysis. The task was concrete: take an earnings call transcript, analyze it, summarize the key points. He compared the model’s output to his own analysis.

The moment of conviction came when he reached indistinguishability. “Once I got to a point where I was fine tuning very early versions of GPT3 to recreate, analyzing or summarizing earnings calls to the point where I couldn’t tell the difference between something the model was writing and what I was writing, it became really clear to me that this technology would be wildly impactful to any knowledge based domain.”

Think about what this means. He wasn’t impressed by the technology in abstract. He was convinced because the output matched what he—an experienced investment analyst—would produce. This gave him unshakeable conviction when others were still skeptical.

The Credibility Problem

But conviction doesn’t equal customers, especially when you’re selling something that sounds like science fiction. Remember: this is pre-ChatGPT. AI hadn’t broken through to mainstream consciousness. The idea that software could be a “thought partner” for institutional investors sounded far-fetched.

“AI as a thought partner on an investment team was something that felt very far fetched,” David acknowledges. So how do you bridge the gap between your conviction and their skepticism?

David had one advantage: “I had been in the industry for a bit, working at a few different funds. So I had some relationships with people that trusted me and I certainly leveraged those relationships early on.”

But notice what he didn’t do. He didn’t just ask people to trust him. He used trust to buy something more valuable: patience for demonstration. “A lot of my kind of credibility in working and having worked with some of these folks or been colleagues with them helped me like kind of bridge that gap of uncertainty and a commercial relationship in a way that kind of would build trust over time.”

The Co-Founder Decision

David’s self-taught ML knowledge was enough to validate the idea. It wasn’t enough to build a company. “I’d say actually the first thing I did was find a tremendous Co-Founder who is a deep technical expert and just a really strong full stack engineer.”

The timing of this decision reveals sophistication. Many non-technical founders find a technical co-founder immediately, before they really understand the problem. Others wait too long, trying to bootstrap too far on their own.

David found the middle path: experiment enough to develop conviction, then find the technical partner who could accelerate execution. “That was super important because it gave me the capability to experiment and iterate quickly to start testing out some of these concepts I had at that point.”

The Experimentation Gauntlet

What followed was months of intense learning. “A lot of it really was some combination of experimentation, both like between my Co-Founder and I, as well as with very early trusted kind of design partners and just understanding the capabilities of these models where they can be useful and how they could impact research workflows.”

Notice the dual track: internal experimentation plus customer validation. They weren’t building in a vacuum, but they also weren’t just taking customer feature requests. The goal was understanding fundamental capabilities—what these models could actually do versus what people hoped they could do.

David was explicit about recognizing his blind spots: “I’m certainly have my own set of biases and whatnot as a potential user of this technology.” So they talked to more users, iterated more, prototyped more. “Yeah, a lot of what were doing there was just speaking with these potential customers, iterating, experimenting, prototyping.”

The Funding Bootstrap

While experimenting, David was also collecting angel checks—small investments from individuals who believed in the vision. This created runway for deeper experimentation without the pressure of a formal seed round with its expectations and timeline.

Eventually, these angel checks led to a pre-seed round. But the sequencing matters: conviction from experiments, initial validation from design partners, angel funding to extend runway, then institutional pre-seed once the foundation was solid.

This is radically different from the raise-first-build-second approach many founders take. David built conviction through experimentation before asking for significant capital.

The Timing Paradox

Building before the hype cycle created both advantages and disadvantages. The disadvantage: everything was harder to explain. Every prospect needed education about what language models could do. Every investor needed convincing that this wasn’t just research curiosity.

But the advantage? When ChatGPT launched in November 2022 and suddenly everyone cared about AI, Portrait had years of head start. They’d already worked through the hard questions about where these models worked and where they didn’t. They’d already built credibility with early customers. They’d already refined their product through iteration.

“We’re a little bit lucky in that. Generative AI is quite topical and a lot of investment firms are quite focused on it. So we’ve gotten a lot of inbound,” David acknowledges. But that luck only mattered because they’d put in the work before the topic became hot.

The First Customer Playbook

Converting early relationships into paying customers required a specific sequence. David broke it down: early prototype demonstrations showing concrete examples of impact, converting those demonstrations into letters of intent, then delivering actual product to convert LOIs into revenue.

“A lot of it was really early prototype demonstrations of what we could build, examples of like how that could impact specific aspects of the idea generation and context building process and yeah, turning those into initially kind of letter of intents for when we had a working MVP and then converting those letter of intents once were able to start, you know, delivering a real product to some of these early users.”

This sequence created commitment in stages. Each step reduced uncertainty while increasing investment from both sides.

The Conviction Compound

Looking back at David’s journey—from Stanford experiments to GPT-3 fine-tuning to design partners to angel checks to institutional funding—the through-line is building conviction through demonstration, not persuasion.

He convinced himself by building something that matched his own output. He convinced his co-founder by showing the experiments. He convinced design partners by demonstrating specific impact. He convinced angels by showing customer validation. He convinced institutional investors by showing traction.

Each step required the previous step’s proof. You can’t skip stages when building something that sounds far-fetched. For founders building in emerging categories before the hype arrives, David’s path offers a template: develop personal conviction through hands-on experimentation, find the technical partner who can accelerate that experimentation, validate with design partners who trust you enough to experiment with you, use early validation to raise initial capital.