Ashish Jha
ML engineer · inference & model optimization · researcher · writer
I'm a machine learning engineer focused on inference optimization, model compression, and data-efficient training, shipping models that run faster, smaller, and cheaper. My research in tensor methods feeds directly into that work. Alongside it, I write and make videos that unpack the ideas behind modern AI from first principles. I also run Frontier ML Nepal, a community for people exploring machine learning in Nepal.
I'm currently open to full-time roles in research or engineering, freelance and consulting work - prototyping, model/inference optimization, post-training, and technical writing or teaching. If you're building something I can help with, get in touch.
Selected Work
mHC-nanoGPT
From-scratch GPT training stack with DeepSeek-V4's Muon optimizer and Newton-Schulz orthogonalization; torch.compile, presets, unit-tested.
GRAFT ★ 6 · paper + code
Cuts training cost ~4x by selecting the most informative 25% of the data, holding accuracy.
Videos
Writing
The Last Conversation
Why a perfect civilization would be a silent one.
The Weakest AI Model You'll Ever Use Again
Claude Fable 5 is the most powerful model ever made public. The unsettling part is what that tells us about next year.
Everyone talks about post-training. Few people explain what it actually looks like.
From instruction tuning to reinforcement learning, a practical look at the systems, datasets, and tradeoffs that turn a base model into a capable coding assistant.
The Most Beautiful Trick in DeepSeek’s V4 Paper Part II
The other side of the coin - residual connections at extreme scale
The Most Beautiful Trick in DeepSeek's V4 Paper
A journey from SVD and geometry to the single operation powering DeepSeek V4's training and architecture.
The Most Beautiful Trick in DeepSeek's V4 Paper - Part I
Part 1: The SVD and Muon