An artificial intelligence assistant is a conversational interface that interacts with users in natural language, understanding their questions and returning useful answers. Unlike traditional chatbots, these assistants can connect to multiple data sources, execute complex processes and adapt to the context of the conversation.

In the field of software development, AI assistants can be key allies in automating repetitive tasks, generating insights in real time, and reducing friction when accessing technical data. Exactly, that is what we are looking for with Pulzen.
‍

Pulzen case: Why did we create an AI assistant?

Pulzen is a platform that measures the productivity of development teams based on real data, such as commits, pull requests, code reviews, DORA metrics, and others. Our mission is to help teams to understand how they work and how to improve without having to interpret hundreds of dashboards.

But we realized something: Even with good data, the experience of exploring it could be much better. We wanted anyone, from a team leader to a developer, to be able to ask Pulzen directly such things as:

“How was the code review time this week?” or

“Which teams lowered their deployment frequency?”,

and get immediate answers, translated into text or graphics.

This is how our AI assistant was born: A productivity co-pilot that understands the context, generates queries, interprets results and guides you with reasoning, not just with data.

‍

Step by step: How we created our assistant with AI

1. Choosing the SDK: Simplified interaction with models

We chose the Vercel AI SDK because it allows us to quickly create modern conversational interfaces. This tool facilitates communication with language models such as GPT-4o and GPT'4o-mini, integrating features such as streaming responses and managing histories.

‍

2. Vector store in MongoDB: Conversational memory

In order for the assistant to remember the context of the conversations, we implemented a history storage system in MongoDB, using vectorization techniques. This allows the model to “remember” previous interactions, offering more consistent and personalized responses.

‍

3. Custom Fine-Tuning: Specific Queries in MongoDB

We train our own models with real examples of queries in our Mongo collections. This allows the assistant to generate optimized queries to answer productivity questions, without the need for the user to know the structure of the database.

‍

4. Thinking steps: Transparency in orchestration

We implement reasoning processes (“thinking steps”) where the assistant shows each stage of the analysis before delivering an answer. In this way, the user can understand how a conclusion was reached and what steps the system executed behind it.

‍

5. Agent-based architecture: Intelligent Routing

We design an architecture of specialized agents. Each agent has a role—for example, analyzing DORA metrics, visualizing graphics, or interpreting text—and the system chooses which one to activate based on the user's intention. This approach based on Routing allows for more precise and flexible answers.

‍

‍

AI assistants are changing the way we interact with information: it's no longer just about searching, but about talking, understanding and deciding with the help of artificial intelligence.

At the Pulzen team, we believe that these types of tools not only save time, but also democratize access to metrics that previously required technical knowledge. In the future, we imagine even more proactive attendees, recommending improvements, detecting critical patterns and helping teams stay focused on what really matters.

This is just the beginning.

‍