Working on the AI layer of an early-stage startup.
Working part-time alongside the CEO and engineering team on AI, data and the product systems around them.
How I got involved
A friend launched a startup in early 2026 and reached out a few months later because he needed someone work on the AI and data side of the product.
What I have been working on
The work sits somewhere between applied AI, backend engineering and product development.
So far, that has included:
- complex LLM-based parsing pipelines;
- self-learning mechanisms built around model feedback and accumulated system context;
- conversational systems that interact with structured business data;
- backend infrastructure around those AI capabilities;
- parts of the user-facing experience needed to make them usable.
One of the more recent pieces is a conversational NLQ agent that can answer questions about business data while interacting with a much more complex underlying system.
The interesting part is not getting the model to answer a question. That part is usually the demo. The real work is deciding how it gets to the data, what it is allowed to see, how the system handles ambiguity, and how to make the whole thing reliable enough that people can actually depend on it.
AI is only one layer
I joined primarily for AI and data, but the work rarely stays there. Shipping an AI capability usually means touching APIs, backend systems, data models and, occasionally, the frontend too.
That has made the role much closer to building product systems around AI than simply integrating models. It is also a good reminder that a very impressive notebook is still quite far away from a feature someone can safely use.
Current status
I’m still working with the team part-time. The company is currently in stealth, which means I can talk about the shape of the work, but not most of the interesting context around it.
What I’m taking from it
AI systems fail at the boundaries
Parsing, retrieval and generation matter, but the harder problems often live between the model, the data and the rest of the application.
Product AI means owning more than the model
The useful work is often backend, permissions, data contracts and UI behavior around the model, not the model call itself.
The demo is not the product
Getting a model to do something impressive once is usually much easier than making the behavior reliable, bounded and useful enough to ship.