Sunny Ray sits down with Saurabh Mishra, founder of CodeFlash, to trace his path from a curious kid in India who took things apart, through an electronics engineering degree at IIT (BHU), to Nvidia, a master's at Carnegie Mellon, and eventually Meta, where he worked on recommendation systems serving over a billion users. Saurabh describes how his lifelong obsession with efficiency, shaped by conversations with his mechanical engineer father, collided with the release of GPT4, which he discovered could meaningfully optimize existing code. That moment convinced him software's hidden crisis was solvable: as AI coding tools generate more code, they quietly produce more slow, unoptimized systems because engineers no longer deeply understand what they ship. He explains why performance has always taken a backseat to feature development and how CodeFlash automates optimization by running in the background on pull requests, searching for the most efficient correct implementation, and surfacing suggestions for review. The conversation covers his engineering formation across hardware and software, his entrepreneurial awakening at CMU, and why he ultimately left Meta to build CodeFlash full time.
An ex-Meta engineer explains why AI is quietly making software slower and how he is fixing it.
Tell people one or two lines on the company you're building, CodeFlash.
I'm building CodeFlash and we are solving the problem of making all generated and human written code always be optimal and performant.
What drew you from electronics engineering into AI and software startups?
I always thought of it as understanding the whole stack, from the physics of chips to computer architecture to programs to applications. Having that formulation made me a much better engineer because I understand everything from the physics level to the system architecture level. Understanding the whole world was always one of my passions.
What was it like working so close to hardware at Nvidia, building performance tools for chip architects?
Everything we run just triggers switches in a chip, so understanding how graphics flows through a GPU and how to optimize it at a whole chip level was fascinating. We worked on bleeding edge GPUs, which threw me into the deep end of performance, since GPUs being the fastest is Nvidia's main thing. We also used AI to improve GPU performance, which was very early but helped me grow quickly.
At Meta you worked on recommendation pipelines for over a billion people. What did that scale feel like, and what cracks did you notice?
At Meta's scale you can form a hypothesis and validate it with experiments almost immediately, which makes every decision very quantitative. I once ran an experiment that raised Instagram's daily active users by 1 percent, which is the kind of thing only possible at that scale. I was running ML pipelines processing data on around 10,000 nodes, which was enormous.
Where did your obsession with efficiency come from?
It's partly from childhood. My father is a mechanical engineer, and talking with him and other engineers made me realize engineering progress happens when you increase efficiency, doing more with less. I felt that efficiency had been missing in software design, and peak efficiency felt like the peak of engineering, which really drew me in.
Millions of engineers see slow code but don't quit to start a company. What made you finally act?
I'd wanted to start a company since CMU in 2018, but inspiration takes time and the timing wasn't right. When GPT4 came out I was playing with it and realized it actually worked well for optimizing code performance. I couldn't sleep for three or four days because I understood this was the missing piece, and after talking to customers and investors I got strong positive signal and took the plunge.
Was there a hackathon that demonstrated LLMs could optimize existing code?
GPT4 dropped on a Thursday while I was stuck on other startup ideas. AGI House was hosting a GPT4 hackathon, so I went and played with GPT4 that Friday evening and realized it could actually optimize the performance of code. Everyone at the hackathon presentation liked it.
Why was the idea of AI improving code so counterintuitive, even to seasoned engineers?
Optimization is essentially the expertise problem, and early on people believed LLMs wrote terrible, hallucination prone code, so the skepticism made sense. But counterintuitively, the amount of slow code is increasing now because teams relying heavily on tools like Copilot or Codex ship things that work without understanding performance, so they end up writing more slow code than an attentive expert would.
Building something daring? Sunny talks to founders like this every day. Fifteen minutes to see if your story belongs on the stage.
Claim your pre-interview