Pratyush Mittal

About Me

I'm a backend engineer who ended up loving the parts of software most people avoid — distributed systems, real-time infra, the stuff that has to work instantly or not at all.

What I'm good at

Give me a system that needs to handle a lot of traffic without falling over, and I'm happy. At KreditBee I built our web calling platform from scratch — WebRTC, SIP, WebSockets, the whole real-time stack. It's live in production now, handling actual calls, and getting it there meant a lot of late nights figuring out why a call would randomly drop or why latency spiked under load. That kind of debugging is honestly where I learn the most.

I've also worked with LLMs in a way that's a bit more hands-on than most people get to. During my internship at Samsung I fine-tuned a model using RLHF and RLAIF, and later at KreditBee I built a Gemini-powered service that automatically reviews call quality and flags compliance issues — something that used to be done manually by a team of people.

What drives me

Honestly, I like fixing things that are almost right. There's a specific kind of satisfaction in taking something that technically works and making it actually reliable — the kind of reliable where you stop thinking about it because it just doesn't break. I don't love sitting around planning for weeks either. I'd rather build something rough, see where it breaks, and fix it from there.

A project I'm genuinely proud of

At Samsung, part of my job was making an LLM less toxic — which sounds simple until you're actually doing it. I ran PPO-based reinforcement learning across a bunch of training runs and got the model's toxicity score down from 2.51 to 1.05. That's more than half.

What stuck with me wasn't the number though, it was realizing how much of alignment work is just careful iteration — training, checking what changed, figuring out what broke in the process, adjusting, repeating. It's not that different from debugging a production system, honestly, just with a much fuzzier definition of "bug."