Where does the waiting happen?
Change the workload and request strategy, then watch an interface get its data.
Siddharth Agarwal · Darmstadt, DE
I’m Siddharth. I spent four years shipping a Jira-integrated SaaS product in Java, TypeScript and Vue.js. Now I build applied and agentic AI while studying Data Science at TU Darmstadt.
Open to Werkstudent roles and internships in Germany01 / Selected experience
Four years building B2B software at Navarambh, and an internship in test automation before that.
Navarambh Software Pvt. Ltd. · India
AIO Tests is a test management app that runs inside Jira, listed on the Atlassian Marketplace. QA teams use it to write test cases (including with AI), run test cycles, track defects and report on testing without leaving Jira.
Experience and metrics from my résumé. Employer source code isn’t published here.
02 / Recent project
Black Forest Hackathon, Oct 2026 · Team of five · Challenge set by KOCH Freiburg GmbH
AI agents that call business tools directly tend to pull whole tables into their context. That is slow, expensive, and often leaks data the agent never needed.
A sandbox for AnythingMCP, an open-source MCP gateway. The agent writes a short JavaScript script, the script runs next to the connectors in an isolated V8 sandbox, and only the final answer goes back to the model.
Understanding the problem. KOCH wanted permissions and different levels of access built in, so we had to find a design that worked well, looked good and fit their setup, while still covering everything they expected.
Keeping data out of the model means running code the model wrote. So the sandbox gets the user’s own permissions and nothing more: it refuses writes, has no network access or secrets, cannot reach other tenants, and stops runaway scripts.
I broke down the challenge and decided with the team what to build, steered the build (much of the code was written with AI coding agents), and debugged and fixed the issues that came up on the way to a working demo.
Agent is asked: “What happened to order 12345?”
const o = await tools.orders.get({ id: 12345 }) const u = await tools.users.get({ id: o.customer_id }) const s = await tools.shipments.find({ order: o.id }) return { status: s.status, eta: s.eta, customer: u.name }
Same answer, about 10 tokens instead of 422,500. From our hackathon benchmark.
03 / The playground
Four things to try, from the request timing behind my backend work to two game bots that explain their moves. No accounts, no setup.
Change the workload and request strategy, then watch an interface get its data.
Tic-tac-toe, solved completely. Connect Four, too big to solve, so its bot estimates. Both show their reasoning.
Draw a maze and watch breadth-first search and A* race for the shortest path.
Ask about my experience, AI work and skills. Prepared answers drawn from my résumé.
Made for this site with help from AI coding tools · Everything runs in your browser · The assistant gives prepared answers from my résumé, it isn’t a language model · Source on GitHub
04 / A little context
I like the part of engineering where a product gets noticeably better for the people using it.
A page that loads faster, a migration nobody has to think about afterwards, an AI feature that saves someone real work instead of just demoing well. That’s what four years on AIO Tests taught me to care about.
Today I’m doing an MSc in Computer Science with a Data Science focus at TU Darmstadt, and exploring applied and agentic AI alongside it. I also love hackathons: an unfamiliar problem, a new team and a deadline measured in days. So far they’ve had me build everything from a gesture-controlled ATM PIN pad to a sandbox for AI agents.
The full story, in one PDF ↓Winning ATM authentication prototype · Smart India Hackathon 2020
First place for NCR Corporation’s problem statement NH447, with an ATM PIN entry system controlled by hand gestures.
TU Darmstadt · Data Science focus
Navarambh Software · AIO Tests for Jira
Frugal Testing · JMeter load tests and Java/Selenium automation
Thapar Institute of Engineering & Technology, Punjab
English (C1) · German (learning, B1)
A working 3D printer set up for remote printing over the network with Raspberry Pi and OctoPi.
An autonomous line-following delivery robot using Arduino, ESP8266 and Firebase.
An Android app that places 3D floor plans on printed markers, built with Unity and Vuforia.
05 / Let’s make something