Hi, I’m
Himanshu Mittal
Currently leading technical work on frontier-model safety and evaluation, especially interested in scalable oversight. Previously built industrial optimization systems and a venture-backed hardware-AI startup.
01 Current / AI safety
Making the model safe is not enough.
Safety also depends on the mechanisms that govern its deployment, and on whether meaningful human oversight still exists.
I use Claude Code daily. It is fantastic. But occasionally I see something unsettling: it fails a test and edits the test until it passes instead of fixing the bug. It provisions resources nobody asked for. It drifts from the original instruction and still reports success.
The result can look right even when the process wasn’t. If I can be misled by a clean-looking result on a task I specified carefully, what happens as we delegate work that is larger, faster, and harder to inspect?
Verification has to become cheaper than the work being verified. I don’t think we know how yet. At Solumn AI, I work on this from the evaluation side: RL environments, adversarial tool-use tasks, and automated red-teaming for frontier models. Much of that work examines cyber harm and failure modes such as deception, misaligned autonomy, and insecure code.
Read: Red Teaming GenAI Safety as a Measurement Problem →
Happy to talk.
02 Selected work
More writing →Faradworks
Three Questions From Building an AI Hardware Startup
Read the story →Optimus
Building Optimus: What It Takes to Schedule a Pharmaceutical Plant
Read the story →From the archive
Explore the archive →Voice-controlled chess 2018
A Python course ended with tic-tac-toe and blackjack. I wanted something with enough logic that I couldn't hold the whole thing in my head at once. Chess seemed like a good idea. Then I decided I should be able to say a move out loud and watch the board respond.
It became the first thing I ever pushed to GitHub. I still play occasionally, around 2000 on Lichess.
Repository ↗TensorRT Swish plugin 2019
An NVIDIA bootcamp gave us 45 days to build a native TensorRT plugin. My teammate and I spent half of them solving the wrong problem, until a session with NVIDIA engineers cleared things up. I must have read the developer guide fifteen times. We wrote a Swish plugin in C++ and measured roughly 3× faster inference on LeNet-5.
We came first. There were two teams.
Repository ↗Crop radiation prediction 2020
Under the guidance of Dr. Gaurav Singhal, I built the training pipeline: regression splines, gradient boosting and a small neural network, all driven by one config file.
The work was published in the International Journal of Biometeorology. My first peer-reviewed paper.
The paper ↗MyGSTConsultant 2024
I saw my father working through a GST calculation in Excel and thought it looked like a ten-minute scripting job. Then he explained all the rules and edge cases hiding underneath it. I built the first calculator in a day, and we have kept adding tools occasionally.
Website ↗Turing 2024 → 2025
After Ripik, I moved to Turing and spent about a year creating and reviewing SFT/RLHF data for large language models. A lot of it was around implicit code execution: coding tasks for data and ML analytics, document generation, model-trajectory comparisons, and classifying behavioural logs against detailed rubrics.
It was probably the most repetitive work I have done, but also the first time I spent months looking at model behaviour one example at a time. That eventually pulled me towards evaluation, agents, and AI safety.