David Feldell

My view has shifted from AI as a better interface to AI as a reasoning layer that can understand how an ecommerce business operates, recommend what should happen next, and automate workflows end to end.

The biggest opportunity is not just improving individual tasks. It is giving users enough freedom to build around how they actually work: agents, applications, workflows, and new ways of interacting with the business.

What surprised me most is how capable people become when the product gives them that freedom. Users do not need every workflow to be predefined for them. They quickly learn how to combine data, intelligence, and agents in ways we would not have designed ourselves.

I have also been surprised by how quickly people have adapted. Behaviours that felt unfamiliar a year ago, like delegating work to agents, building internal tools through conversation, or trusting AI to take several steps autonomously, are becoming normal much faster than I expected.

Dema is becoming the operating system for AI-native ecommerce companies.

Customers can already build agents, applications, and workflows on top of a unified intelligence layer in Dema. Over the next year, that system will become more proactive, more personal, and more deeply embedded in how companies make and execute decisions.

The goal is to collapse the distance between understanding what is happening, deciding what to do, and taking action. Today, those steps are spread across dashboards, spreadsheets, meetings, and disconnected tools. Dema brings them into one intelligent system.

What makes this powerful is the compounding loop underneath it. Dema combines real-time commerce data, predictive and causal intelligence, company memory, agent behaviour, and application usage. As the system is used, it learns more about how the company operates and becomes increasingly valuable.

Most people are not looking for another transformation project. They want less manual work, better decisions, fewer stressful follow-ups, and more leverage.

That is how Dema becomes the company brain: not as another tool employees need to manage, but as the system through which more of the company increasingly operates.

Our biggest bet was that strong AI products would need to own much more than the model.

Ecommerce data is fragmented across behavioural events, orders, advertising platforms, inventory, deliveries, returns, and financial systems. Collecting it is not enough. It needs to be transformed into a unified model where everything can be joined and understood accurately.

We began building that foundation before the current generation of LLMs was capable enough to fully use it. Two years ago, we were already structuring data to be accessible to future agents, not only dashboards and analysts.

We also enrich that foundation with predictive and causal intelligence, including forecasting, marketing mix modelling, incrementality testing, attribution, and customer lifetime value. The goal is not only to understand what happened, but why it happened, what is likely to happen next, and which actions will create the most value.

But owning the data and intelligence layer alone is not enough. We made a deliberate choice to also own the agent harness and the application layer.

The data provides context. The predictive and causal models create deeper intelligence. The harness makes agents reliable. The application layer turns all of it into real outcomes for the customer.

I try to create a culture with a very high level of product ownership.

We do not have product owners handing engineers a predefined roadmap. Engineers are expected to understand the customer, the product, and where the company is heading, and then help decide what should be built next.

I believe that great people should have a lot of freedom. If you have strong judgment, understand what matters, and can connect your work to customer and business value, you should not need someone constantly telling you what to do. The best ideas often come from the people closest to the problem.

We support that with short cycles, early releases, and close iteration with customers. The underlying data needs to be trustworthy, but the product does not need to be perfect before users can start getting value from it.

As AI allows people to work across more of the stack, that freedom becomes even more powerful. Engineers, designers, and product-minded people can solve problems beyond traditional role boundaries, which makes trust, communication, and shared context increasingly important.

I also care deeply about building teams where people genuinely enjoy working together. People do their best work when they care about the problem, trust the people around them, and feel real ownership of the outcome.

We try to balance ambitious, curious people from many different backgrounds with senior engineers who bring experience, judgment, and architectural depth.

Curiosity and problem-solving ability matter more to me than a specific background.

The strongest people are not satisfied with simply doing what they are told. When they encounter a problem, they investigate it, understand the root cause, and consider several different ways to solve it.

Most problems can be solved in ten different ways. What matters is whether someone can navigate that ambiguity and make a thoughtful decision.

I look for people who have developed their own way of becoming good at new things. They know how to learn, experiment, ask good questions, and build understanding in an unfamiliar area.

We are already at the point where a customer issue or product idea can move from Slack to production with agents doing much of the investigation and implementation.

The work is now moving to a higher level of abstraction.

Engineers spend less time manually writing every line of code and more time designing the systems that produce good outcomes: architecture, context, interfaces, constraints, evaluations, and feedback loops.

We have moved from mainly reviewing pull requests to increasingly reviewing plans and ADRs. The important questions are becoming whether the problem is framed correctly, whether the architecture is sound, and whether the agent has enough context to execute well.

The same shift is happening in design. Designers can increasingly move directly into frontend implementation rather than stopping at static designs.

As execution becomes cheaper, judgment, taste, and systems thinking become more valuable.

The hardest problem has been building a real-time intelligence platform that unifies behavioural and operational data across many different ecommerce systems, while also supporting the realities of onboarding very different customers.

We process high-volume behavioural events in real time and turn them into sessions, journeys, and customer-level intelligence. That means handling late and out-of-order events, identity resolution, and continuously changing state.

At the same time, we transform data from advertising platforms, ecommerce systems, inventory tools, delivery providers, and return systems. Every source has different schemas, identifiers, definitions, update frequencies, and levels of granularity.

Then we need to make all of that work across customers with different technology stacks, catalogues, commercial models, and data quality.

The challenge is balancing standardisation with flexibility: building one unified intelligence layer without creating a separate platform for every customer.

It matters because predictions, recommendations, workflows, and agents are only as strong as the context beneath them. To automate meaningful decisions, the system needs data that is timely, accurate, connected, and trusted.

Outside work, I enjoy running and spending time with my family. Running helps me clear my head and create some distance from the constant pace of building a company, while time with my family helps me stay grounded and present. Both give me the space to reset, and often that is when I return to work with the clearest perspective.