Petras Bukelis

Seven apps, one shared foundation

Seven standalone applications — spanning data pipelines and language-driven analysis — each its own codebase, each operable only by a developer, each rebuilding the same groundwork from scratch. I architected a shared services layer underneath all seven, plus a management interface so non-developer staff can configure and run them directly.

Led the architecture end to end — my first project of this scale — refining the design with feedback from our CTO along the way.

Don't hand a stranger your personal data

When large language models started getting wired into internal products, there was no way to guarantee a third-party provider wouldn't retain — or leak — the personal data inside the text it received: names, phone numbers, ID codes, addresses. Rather than avoid LLMs, I built a self-hosted service — running its own local NLP model — that strips personal data out before any of it reaches an external API.

Two things stood in the way: I didn't know Python, and I'd never administered a production server. I used LLMs to get up to speed on both fast enough to design, ship, and still personally run the infrastructure it lives on.

It cross-checks text against roughly 8 million public first and last names across six languages, catching an estimated 94–95% of personal data across 10M+ tested examples, with a false-positive rate around 2–3% — reviewed and tightened every month against real production logs. It's been running in production, self-hosted and self-managed, for about two and a half years without a security incident.

When the client asked for something that didn't exist yet

A client wanted their own non-technical staff to ask questions about their survey data directly — no developer, no dashboard, no waiting on a report. At the time, nothing packaged covered it: Claude Desktop's connector ecosystem was brand new, and I'd never used the MCP protocol before. About a day of research later, I had an architecture that worked.

The result: staff type a question into Claude and get an answer straight from the live data — installed on their own computer, Windows or Mac, with no developer in the loop.

That gap hasn't closed since. The database we use still has no official connector in Claude's ecosystem — some other databases do now, but none of them hand you the same level of control over exactly what a connection can and can't touch. Simpler to install isn't the same as fully solved.