Loopin

An AI agent that remembers what your whole team did — so you never have to ask.

Comment

What's next for Loopin

Here's a full Project Story in the standard Devpost format (Inspiration → What it does → How we built it → Challenges → Accomplishments → What's next). You can trim once you've actually built it and know what really happened.

Inspiration

Every small team wastes hours on the same ritual — "what has everyone done so far?" Standups, status-check pings, and "quick updates" interrupt focus and rarely give a complete picture, especially in async teams where work happens at different times. We wanted an agent that removes this friction entirely: instead of people reporting to each other, the team's actual work — commits, messages, task updates — becomes a queryable memory anyone can tap into instantly.

What it does

TeamPulse passively ingests activity from a team's tools (starting with GitHub commits), summarizes and structures it using Gemini, and stores it as a persistent, queryable memory. Instead of pinging a teammate, anyone can ask the agent directly — "What has X done on the auth feature this week?" or "What's blocking the launch?" — and get an accurate, synthesized answer in seconds, with zero manual reporting from the team.

How we built it

We built TeamPulse using Gemini 3.5 for both summarization and question-answering, orchestrated through Google's Agent Development Kit (ADK). Activity data is ingested via the GitHub API, condensed into structured entries (who, what, when, why) by Gemini, and stored in Firestore as the agent's long-term memory layer. The whole system is deployed on Cloud Run, so ingestion and querying both run as a live, always-on service rather than a local script.

Challenges we ran into

Turning raw, messy activity (commit messages, PR descriptions) into structured, queryable memory without losing important context was the hardest part — we had to carefully prompt Gemini to summarize consistently rather than just paraphrasing loosely. Designing a Firestore schema that stayed simple but could answer both specific ("what did X do") and broad ("what's blocking launch") questions accurately also took several iterations.

Accomplishments we're proud of

Getting a fully autonomous pipeline working — from raw commit to structured memory to natural-language answer — with no manual data entry required. It genuinely replaces the "checking in" ritual with a system that just knows.

What we learned

We learned a lot about designing agent memory that's both structured enough to query reliably and flexible enough to handle open-ended questions — plus practical lessons in deploying agentic workflows on Google Cloud with ADK and Firestore.

What's next for TeamPulse

Expanding ingestion to Slack and task trackers (Jira/Linear), adding proactive alerts (e.g. "X hasn't touched this in 5 days"), and building a lightweight dashboard view alongside the conversational agent.

Built With

Share this project:

Updates

Submission history