Balvinder
Singh

A curious software engineer

Rome could be built in a day with Claude.

Based in Paris. Looking for end-of-study internships (2027): AI, software or infrastructure engineering.

Right now

Learning how the models work

I’m doing an M2 in Artificial Intelligence at CentraleSupélec, in the track on AI for societal challenges (IA appliquée aux enjeux sociétaux).

Next: a six-month end-of-studies internship in 2027, in software, infrastructure or machine learning engineering.

Things I’ve built

  • Pulse

    A traffic layer in Go. Load balancer, Redis-backed rate limiter and cache. Instrumented with Prometheus, then pushed with k6 to see where it bends. It bent at the load balancer, not the replicas.

    Read the dig notesCode

  • AgentAudit

    An LLM agent that answers portfolio-risk questions by calling Java tools, with every tool call, argument and result kept in PostgreSQL. The record caught the model saying EUR 12M was more than EUR 14M.

    Read the dig notesCode

  • Medical AI research

    Research on applying AI to medicine.

    Details on my resume

2025–2026

Keeping OYO running

I joined OYO’s Merchant team as an intern in January 2025 and stayed on as an SDE until September 2026.

What shipped: backend APIs for smart-lock and room-upsell products in the UK and UAE, a self-service digital check-in flow, and dashboards for the Motel 6 management system.

Most of my work sits below the surface: Docker and Kubernetes, keeping services stable, tuning their resources and deployments, and reading logs until a bug gives itself up. Azure for databases.

Where it started

Delhi, then Mumbai

Computer Science at Netaji Subhas University of Technology, specialising in AI and machine learning.

My first internship: software development at Mahindra Logistics in Mumbai, February to April 2024.

Bedrock

What I keep coming back to

Understanding how models work, and building the systems that let them run.

Away from the keyboard: photography (looking for sponsors), F1 (du du du du), lifting (weights and occasionally my morale), and music on in the background (turn up the volume!).

Questions I’m chasing

  • What do you have to record about an AI agent’s run to defend its answer later?
  • Where does a small distributed system break first, and how many runs does it take to be sure?