Amplitude AI Builders: Adam Struck Discusses Moda

This post is part of our Amplitude AI Builder series. Each one features an Amplitude engineer discussing an AI product that they are building.

Amplitude makes a digital analytics platform. We’re building some powerful AI tools that we think our customers can use to be more productive. Except Adam Struck, senior AI engineer at Amplitude. He makes an AI tool for Amplitude employees to use so they can be more productive.

For this post, I talked to Adam about how he and his team built Moda, a powerful AI assistant for Amplitude employees. We covered the origin of the project, why it’s called Moda, how he uses AI to make Moda better, and more.

What is Moda?

Moda is Amplitude's internal AI assistant—it’s designed to make finding and understanding information super easy for employees. It uses Glean to search across all sorts of platforms like Confluence, Jira, Salesforce, Google Drive, Zendesk, and Slack. Plus, Moda ties into Amplitude's MCP server and other custom internal tools for more advanced use cases.

Most people use Moda in Slack, since that's where they spend a very large amount of their time. It can be used like a chat bot, but it’s more than that. In Moda’s internal web UI, it has more advanced functionalities, like multi-step workflows to write PRD documents or NCTs.

How did the Moda project get started?

Moda spun out of an AI Week Amplitude hosted back in June. The MVP was created by a talented cross‑functional crew including Eric Carlson, Boogie Cigdemoglu, Nirmal Utwani, and Arav Kumar. I was hired shortly after AI Week to lead internal AI tool development and took ownership of Moda.

From the outset, the focus was on iterative search/refinement, transparent citations and confidence, and tight access controls. As quality improved and usage grew beyond product and design into go-to-market, the project evolved from a prototype into a company-wide assistant with expanding tools and workflows.

Why is it called Moda?

Amplitude’s mascot is the Data Monster. The name “Moda” is a playful reversal and shortening of “Data Monster.”

What’s your building process look like?

It’s probably no surprise, but I use AI tools all the time. I’m especially impressed that I can use Moda to learn more about Moda itself—that kind of meta analysis is super helpful! I can ask how Moda’s being used or where it's not quite hitting the mark. Then, Moda helps me sort and prioritize my feature backlog. When it comes to the actual coding, I prefer using Cursor.

What kind of direction do you give Moda in the system prompt?

I’ve spent a lot of time refining the system prompt, which is essentially the playbook that tells Moda how to think, what to search, and how to adapt. For each search path—whether it’s Glean, the web, or Salesforce—I’ve designed decision trees that dictate which tools to use, how to iterate on results, and when to pause and reflect. That structure makes it possible to refine queries step by step and evaluate answers in a consistent way, which is critical if you want high-quality outputs at scale.

I also put a lot of emphasis on response standards. The prompt enforces structure, tone, and terminology, and it’s very explicit about where citations go and what metadata needs to be included. That makes answers both reliable and easy to audit, while also minimizing hallucinations.

Beyond the prompt itself, I design the tools Moda can call and train it to use them properly. That combination—rigorous prompting plus purpose-built tools—lets me shape not just how the model reasons, but also the kinds of answers it’s able to produce.

Right now, the system prompt is about 350 lines. It encodes rules for assigning confidence levels, checking source recency, and knowing when to push for more evidence. All of that helps Moda clear a high-confidence bar and deliver answers that people can actually trust.

What do people tell you about how they use Moda?

What I hear most is, “How do I get more out of Moda?” Sometimes that’s the entire focus of a meeting. The feedback is enthusiastic but also very practical—people ask, “How would I use it for X?” That tells me Moda is delivering real value, not just novelty.

I keep a close eye on the public Slack channel, jump into threads with tips, and adjust the system prompt or tools when I see an answer that could be better. I also maintain an archive of real questions and conversations, which has been invaluable for tuning and regression testing. Beyond that, Moda has an official feedback mechanism built into every response, so users can easily share what’s working and what’s not.

And I get plenty of DMs—requests for new data sources, new capabilities, or guidance on specific workflows. Those inputs directly shape the roadmap, helping me decide what to connect next and where to refine things so Moda keeps getting more useful where it matters most.

What’s the long-term vision for Moda?

The long-term vision is for Moda to become an end-to-end assistant that can take an idea all the way to execution. Imagine someone starting with market analysis, then moving seamlessly into drafting a product requirement doc, generating mockups, writing code, and even creating sales materials—all within one flow. The goal is for Moda to guide that journey, step by step.

Equally important, I don’t want Moda to be just another tool in a crowded AI toolbox. Inside Amplitude, it should be the trusted assistant—fast, accurate, and reliable—so the team knows they can turn to Moda to get exactly what they need done.