Eashan Babber

Why robotics

I didn't know anything about robotics before coming to UCSC, so there is no robotics on this site yet. That is also why I chose Robotics Engineering: it interests me, and I want to learn it.

How I use AI

I don't use AI for schoolwork. Doing that work myself matters to me, because the more I understand on my own, the better my ideas will be. Outside of schoolwork, I use AI to learn, as a teacher that explains things instead of giving me the answers, and to bring my ideas to life.

I also built this site with AI, to show how I work with it.

What I'm into

I love diving deep into topics that interest me. Right now those are agentic development, game development, mobile development, networks, and my artistic passion, photography.

This room

This site is a listening room. The record holds my projects and open-source work, and the things around the room are the rest of me, the way they would be in a real listening room.

Why I built it

His usage ran out mid-task, and by the time it reset, Claude had forgotten everything. My friend, who later built the Codex version, thought a lot of people were probably frustrated by this, and we spent a week researching how to solve it.

What I built

I built it for Claude Code, and my friend built it for Codex, OpenAI's coding agent, because those were the agents we used. The hardest design choice was what the agent should do when little usage is left: stop, or warn the user and let them choose whether to continue at full scope, switch to a cheaper model, or pause? We chose to warn and let the user choose, because it gives them more control.

Where it is now

We've only tested it on our own devices and agents, so we don't have user feedback yet. The screenshot above shows a real example of the agent's behavior changing.

Why I shared it

I built it mainly for myself. Once I figured out I could package it as a Claude Code mod, a plugin anyone can install, I made it public.

How it works

It turns diagrams written in Mermaid and PlantUML, two text-based diagram languages, into real pictures in terminals that support Kitty graphics, a way for a terminal to show images. The problem was that only some terminals support it, so I made the mod detect the terminal and fall back to the next best option, a diagram drawn in text, when it doesn't. I've only tested it in the Rio terminal on Windows, and a diagram takes a few seconds to load.

What's next

So far it's one mod, tui-diagrams. I'll add more as I find things I want my terminal to do.

What it is

I didn't create GBrain; it's Garry Tan's open-source project. What's mine is how I set it up and run it for my agents, and the fixes I've contributed back to it, listed under GBrain upstream.

How I got here

I started with CLAUDE.md and AGENTS.md, Markdown files that agents read as memory.

Then I tried claude-mem, an early tool for agent memory. But as projects got bigger, the agents' context filled up with project memory, and they would only sometimes update those files. Then I found GBrain, which links Markdown files through a vector database so agents can search them by meaning.

What I run now

I've been building up my GBrain since June 2026. I also set up a separate business GBrain that other people can connect their own agents to, and weekly GBrain health reports that arrive on my phone through Telegram.

A real example

In the screenshots below, I simply say “time to do lecture 10 cse 20.” It knows exactly what I mean from earlier lectures and how I like to be taught, for example, not giving me the answers to the lecture challenges.

What I built

I set up link tracking in the backend so we can see which channels users arrive from. That lives in an admin panel I can't show. Then I built rewards on top of those links, which became the Bring a Friend feature in the screenshot.

Working with agents

Agents often say confidently that they know a library's documentation from training, even when it has changed. What has worked for me is connecting them to current documentation through MCP servers, a standard way to give agents tools and data. Google's developer documentation server, for example, helped me implement Google sign-in in the GoldTickets app.

Beyond engineering

Since playDTC is a startup, I take on more than engineering. I manage our infrastructure: the Neon database, Railway hosting, and the GitHub organization.

I set up a tailnet, a private network between my own devices made with Tailscale, so I can reach my home PC from UCSC without exposing an SSH port to the internet. It also lets me move files between my devices easily.

I originally thought Scouting was going to be weird and not worth it. I was wrong: it's where I learned most of my leadership skills.

As a patrol leader, I led my Scouts through the requirements for the Electricity merit badge, one of the harder badges to earn. I learned how to get younger Scouts' attention and explain things in a fun, engaging way.

I play goalie. Over the past few years I've learned that whoever the captain is, I'm the one directing the team in the pool, because from the goal I can see the whole pool and have to communicate clearly. I play because I care about my health and like being part of a team.

Why a listening room

I have a lot of hobbies, interests and ideas, and together they're what my life is about. I wanted something I can keep adding to that still looks good.

How I built it

I built it with Claude Code and Codex, over many iterations and tests. I gave the agents a reference image for the room and reviewed every version: an early one had objects clipping through each other, and the note cards next to the photos used to be much bigger. Staying involved in the work with the agents is how I got the experience I wanted visitors to have.