When we were sourcing the investments that became Fund I, the technical co-founder profile we were looking for was fairly legible. An NLP researcher or ML engineer, probably with a graduate degree, probably coming from academia or a research-oriented lab at a large technology company, with deep knowledge of specific model architectures and a background in one or two specific application areas. That profile still exists and still produces strong founders. But it has expanded considerably, and the signal that matters most has shifted.
The generative AI era has created a new category of technical founder that did not exist three years ago: the builder-operator who is not primarily a researcher, does not necessarily have a graduate degree in machine learning, and whose technical depth is in applied model integration, fine-tuning methodology, and production system architecture rather than in model research. These founders are proliferating. Some of them are building remarkable companies. The challenge for us as investors is distinguishing the ones with genuine technical depth from the ones who have become very good at demonstrating technical fluency without necessarily having the foundational understanding to build durable products.
Why the traditional signal set is insufficient
The traditional signals we used to evaluate technical founders — graduate degree in relevant field, research publication record, previous employment at a well-known ML research organization — are still relevant but they are no longer sufficient screening criteria in either direction. Not having a PhD is not a disqualifier. Having a strong publication record does not predict product-building ability in the current environment.
The reason the old signals are insufficient is that the technical capability required to build a competitive AI marketing product today is different from what it was in 2021. In 2021, building a language model fine-tuned for marketing copy was a research-adjacent problem. You needed someone who understood the model architecture deeply enough to build and train the fine-tune from a relatively limited starting point. Today, building a marketing-specific model means starting from a foundation model (GPT-4, Claude, Llama, or a specialized alternative), fine-tuning it with proprietary data, building the evaluation and feedback infrastructure that makes the fine-tune valuable, and deploying it in a production system that handles real customer load with acceptable latency and cost. That is more of a systems engineering and product architecture problem than a research problem.
The technical founders who are building competitive products in AI MarTech today typically have strong systems engineering judgment — they know how to design the data pipeline that feeds the fine-tuning, how to build the evaluation framework that tells them whether the fine-tune is actually better, and how to architect the production system for the specific latency and cost constraints of their use case. These skills are not always legible in a resume or a publication record. They show up in how the founder talks about what they have already built and what they learned from building it.
Where the best technical AI MarTech founders are coming from
Looking across our portfolio and the companies we have evaluated but not backed, we see three talent pipelines that are producing strong technical founders in AI MarTech right now.
The first is former founding engineers at AI companies that have been acquired or have scaled. Rui joined Telhaverde from exactly this background — founding engineer at a content intelligence startup that was acquired in 2023. People who built AI products from close to zero, through the growing pains of scaling to real customer load, have a specific practical knowledge that is hard to acquire any other way. They know what breaks in production, they know what the real technical constraints are, and they know the gap between a research-grade system and a production-grade one. That experiential knowledge is the raw material for founding-team technical judgment.
The second is ML engineers from performance marketing and ad-tech companies. The algorithmic advertising infrastructure that major platforms built over the past decade was, in many ways, a training ground for the kind of production ML engineering that AI MarTech requires. Engineers who built and operated ML systems at that scale — ranking models, personalization engines, bidding algorithms — have production-grade ML intuitions that are directly applicable to building AI marketing products. They also have domain knowledge of the performance marketing ecosystem that is relevant to the buyer personas they will be selling to.
The third is the smaller but growing cohort of computational linguists and NLP researchers who have made the transition from academic research to product engineering. The best ones have built a bridge between research depth and product instincts — they understand the models well enough to make informed architecture decisions, and they have enough product experience to know what problems are worth solving. This profile is producing some of the strongest technical founders we see in the content intelligence and language model evaluation spaces specifically.
The questions we ask that reveal depth versus fluency
In a technical founder evaluation, we distinguish between fluency and depth through a specific line of questioning. Fluency — the ability to describe the landscape of current models and techniques accurately — is table stakes. Depth is revealed in how a founder thinks about the specific tradeoffs they made in building what they have built.
We ask founders to describe the worst technical decision they made in the product so far and what they would do differently. This question is not designed to find failure — it is designed to find learning. A founder who can describe a specific bad architectural decision, explain why it seemed right at the time, describe the consequences when it failed, and articulate what they would do differently has a level of technical self-awareness that predicts good judgment in future decisions. A founder who cannot describe a bad technical decision typically either has not built anything real yet, or has not reflected on what they built.
We also ask founders to describe the specific data infrastructure they have built for model improvement and why they made those choices. A founder building on top of foundation models with a genuine proprietary data advantage typically has a detailed, specific answer to this question — specific decisions about what data to capture, how to label it, how to use it in the feedback loop. A founder whose data infrastructure is underbuilt typically gives a general answer about the importance of data quality without being able to describe the specific implementation decisions. The specificity of the answer is the signal.
What we are not looking for
We are not looking for the founder who can most fluently discuss the current state of the art in language model research. That ability correlates poorly with product-building success in our experience. The founders who are most up-to-date on what came out of research labs last month are sometimes the ones most distracted by the research frontier and least focused on building the thing that is in front of them.
We are also not looking for the founder who frames every technical decision as an AI decision. Some of the most important technical decisions in building AI marketing products are infrastructure decisions, not AI decisions. The data pipeline, the feedback loop architecture, the latency optimization — these are engineering problems with well-established solutions. A founder who treats them as AI problems to be solved with the latest techniques is over-indexing on the exciting part of the stack and under-investing in the foundational part. The best technical founders have a clear sense of which problems require genuine research-level thinking and which problems require good engineering judgment applied to known patterns.