Meet Anna
How We Built an AI Assistant That Talks to Enterprise Tools
Part 1 of the “About Anna” series
Enterprise platforms like Pega and Mendix are powerful, but getting anything done in them means you log into a portal, click through a few dashboards, try to remember which button does what, and by the time you’ve context-switched between five different tools it’s already time for lunch. We wanted a way to just open a Teams chat, type a question in plain language, and get a real answer backed by real data from those systems. No portal hopping, no learning curve.
That’s Anna.
Where the idea came from
I’m Ezra, a software engineer at Anamata, a Quality engineering and testing company focused on Modern Code, like Pega & Mendix. My background is a bit different from the rest of the team, I studied Computer Science at the VU Amsterdam, where I got deep into things like compiler construction, operating systems, and network programming. At Anamata, that technical depth turned out to be surprisingly valuable in the projects surrounding our core platforms: I’d write automation scripts, set up containerized environments, build out integrations the platforms don’t offer out of the box.
At the start of 2025, we had an internal session where we reflected on the value we’d added over the past year and where we could add more. I suggested we formalize a “High-coder Team”, a small group with deeper CS backgrounds that could tackle the technical challenges around our Pega and Mendix work. Our CTO Bas and CEO Peter were enthusiastic about the idea, and we started brainstorming what that could look like. Twelve concrete topics came out of conversations I had with colleagues about what could be valuable to offer, things like infrastructure automation, CI/CD pipelines, security engineering, AI-enhanced operations and more. A lot of possible directions for a team that didn’t even exist yet.
The number one priority? Connecting enterprise tools to AI assistants. Basically: give our people (and potentially our clients) a way to talk to their systems in natural language, powered by AI that has access to the tools it’s talking about, instead of a chatbot guessing from documentation.
What Anna does
Anna (short for Anamata Neural Network Assistant) is a Microsoft Teams bot powered by Claude AI. She just lives in the chat environment our team already uses every day, so there’s nothing new to install or learn.
What makes her different from a generic chatbot is that she’s connected to real tools through MCP (Model Context Protocol), a standard that lets AI models interact with external systems in a structured way. I’ll get into how MCP works in a later post, but the short version: when you ask Anna about a Pega case, she actually queries the Pega system and gives you the real answer. Not a summary from documentation. The actual case data. Ask her to book time in our internal time tracking system, and she does it. Ask about a company policy, and she pulls it from the company’s employee handbook.
Anna also knows who she’s talking to. She tracks your role, department, and preferences across conversations, can set reminders for you, summarize past conversations, and will nudge you when you haven’t logged your hours for the week. She’ll reach out on her own when something needs your attention. That comes with privacy considerations, which we’ll address later in this series.
Built on a Raspberry Pi during Christmas
Anna started on a Raspberry Pi 5, sitting in my home office, in the last two weeks of December 2025. My client had wound down for the holidays, so I had two free weeks and an excuse to build something.
I’d been developing on the Pi for a while at that point, it was always on, reachable through my Tailscale network, and I could SSH into it from anywhere, even my phone with Termux. Having the code live on my own infrastructure meant I could just work on it whenever inspiration struck, without worrying about restrictive file-sharing guidelines from client environments.
I used Claude Code as my development partner from day one. I’d never built a Teams bot before, and the Bot Framework documentation is pretty terrible, like it was written for a different SDK and just never updated. Claude Code filled in the gaps though, and the initial scaffold came together in a single evening session: a working bot with dual Claude backends, user profiles, conversation history, and slash commands. 23 files, over 2,000 lines. By 9 PM that same night the project had a name: Anna.
Two and a half hours after the first commit, I was testing her with “What is 2 + 2?” and “What is my name?”, basic stuff, but she answered both. The next morning she told her first joke to a test user in Teams. The responses took a few seconds each and the whole thing was held together with tape. But I kept going.
After that first weekend it snowballed. Anna learned to stream her responses mid-sentence, then to handle group chats where three people talked over each other. After that she started remembering things about you across conversations. All of it running on a Pi 5 that cost less than a dinner for two (well, maybe a bit over because I like tech and spent extra on a flashy case and NVMe storage add-ons). When I showed it to Bas and Peter, Bas’s reaction was basically “wait, this all actually works?” Bas started sketching out client use cases we hadn’t even planned for as well.
What’s next
Next post I’ll get into the architecture, and what it’s actually like building a production system with Claude Code as your daily collaborator. After that, the MCP integrations that connect Anna to Pega and Mendix, what we learned shipping it, and eventually the parts that don’t make it into the demo.
