Founding Soda's AI data-observability research team and product range

Soda builds tooling that helps data teams catch problems in their data before those problems reach dashboards, models, or customers. When I joined — as part of the founding team — that checking was manual and rules-based: someone had to know what could go wrong and hand-write a rule to catch it. That works until it doesn't. It doesn't scale to large, changing datasets, and it only ever catches the problems you already thought of.

For Soda to grow from a popular open-source project into a sustainable commercial product, it needed observability that was automated and intelligent — surfacing the issues users didn't know to look for, not just the ones they'd written rules for.

I founded Soda's Intelligence & Automation team from scratch and grew it into a self-directed group of full-stack data scientists that owned its problem space end to end — from finding user pain to shipping production features.

We ran it in a Shape Up rhythm rather than as an open-ended research lab. Problems were shaped with leadership and straight from customer calls, then given a fixed appetite: a one-day brainstorm to frame a big new capability, followed by two-week cycles of research and prototyping. Promising bets moved toward production; the rest were dropped without ceremony. That kept a research team commercially focused.

When it came to shipping, we didn't throw specs over the wall — the team embedded directly with cloud engineering in a tiger-team model, data scientists and engineers building the production feature together. I acted as the product manager for these capabilities inside Product & Engineering, working daily with 6–8 engineers across backend, frontend, and DevOps, and I built several of the first features end to end myself, from prototype to production.

Over four years the work produced Soda's first suite of AI-driven observability: SodaGPT (an LLM-powered data-quality assistant), time-series anomaly detection, automated root-cause analysis, semantic data-type detection, and automated check suggestions — on a pragmatic Python / Hugging Face / time-series stack.

These features became Soda's strongest lever for converting open-source users into paying customers, and a recurring reason deals closed — they showed up again and again in sales conversations, and helped move the company onto a more sustainable commercial footing at a critical stage.

Along the way I grew the team from zero to four and moved from an individual contributor into strategic leadership and senior product management, helping shape the company's early product direction alongside the leadership team.