Anshuman Mishra
I build AI systems for a living and tinker with them for fun.
At Zomato, I work on evals, post-training, and the infrastructure around both. What started as a bunch of scripts I wrote to solve my own eval problems turned into Gavel - an internal platform I pitched to the VP, got approved, and now serves both AI teams and ops teams across the company.
These days, we are building in-house infrastructure in JAX to post-train small models for real production problems. I am betting pretty heavily on a simple idea: with the right data, evals, and systems work, small models can do far more than we give them credit for - and often replace models that are much bigger and much more expensive.
Before Zomato, I consulted with Google on KerasHub, integrating model families like Qwen and Mixtral into the framework, and worked with Weights & Biases.
I got into ML through open source. My first PR was a small fix to DeepMind’s Optax library — I didn’t really know what I was doing. That led to a Google Summer of Code with TensorFlow, where I ended up shipping GPT-NeoX and Llama 2 for KerasNLP, and eventually co-authoring a paper in JMLR. Somewhere along the way, Google made me a Developer Expert in Machine Learning.
After GSoC I gave talks in 11 cities across India. I still speak at conferences when I have something worth saying.
GitHub | X | LinkedIn | Scholar | Newsletter