hi, i'm rohan.
i build things with machine learning.
mostly interested in:
- neural networks
- training systems
- edge / embedded AI
- ML infrastructure
- making GPUs slightly unhappy
python • pytorch • numpy • rust • c++ • tensorflow/tflite • cuda • occasionally ONNX
esp32-s3 • jetson orin nano
deploying models where they probably shouldn't run
(GRUs and other sequence models on microcontrollers, LLM inference on edge devices)
llama.cpp • ARM builds • edge inference • containerized ML environments
aws • azure
- published work in the Nature Scientific Reports portfolio
- patent granted by the Indian Patent Office
linux mostly.
building deep models for classification, regression, and autoregressive tasks.
comfortable with the full pipeline: architecture design, training, finetuning, and debugging models that refuse to converge.
particularly interested in the math underneath it all: optimization, loss landscapes, gradient dynamics, and why certain models behave the way they do.
LLM security and alignment mechanics.
recently exploring guardrail removal and alignment behavior in transformer models using orthogonal projection–based finetuning: modifying attention up-projections and MLP down-projections to isolate and manipulate refusal directions in activation space. similar ideas appear in tools like Heretic, where safety behavior can be ablated while preserving most of the model’s underlying capability with KL-Divergence very close to 0.
mostly interested in what this reveals about the geometry of alignment in large models.
- training models is controlled chaos
- gradients usually know more than i do
- edge ML is convincing tiny chips to do unreasonable things
- if it runs overnight without crashing, we celebrate
links:
- linkedin: https://www.linkedin.com/in/rsqm/
- github: https://github.com/RsqM