VS Navneet Kanna

About

I got into ML because I wanted to solve hard problems with AI. Once I started learning, I got pulled into how it all actually works under the hood — that’s how dlgrad came about. That’s mostly how I work now: read the papers, build it from scratch, understand what’s really going on.

Education — MSc in Robotics & AI from the University of Glasgow (2023).
Industry Experience — Architected and deployed production-scale Computer Vision systems, building the GPU inference infrastructure that secured a seven-figure (INR) commercial contract.
Projects — Built dlgrad (a from-scratch autograd engine with a C/Metal backend), gradtop (a Rust-based neural network TUI), and several others — see below.
Academic Research — First-author publication, Finding Exoplanets Using Object Detection (Astrophysics & Space Science, Springer 2023).
Technical Writing — Authored a blog series breaking down deep learning concepts, bridging the gap between mathematical theory and hardware execution.

Projects

dlgrad

Deep Learning Autograd Engine

Built a reverse-mode automatic differentiation engine with a PyTorch-like API, developed entirely from scratch to understand deep learning and backpropagation from first principles.
Written in Python with a custom C backend, with no dependencies except CFFI for Python–C communication.
Supports both CPU and Apple Metal GPU acceleration, with autogenerated handwritten kernels optimised for input shapes and operations.
Implemented the full tensor stack: broadcasting, core arithmetic ops (add, sub, mul, div), activations (ReLU), and loss functions (log-softmax, cross-entropy).
Trains networks end-to-end — including MNIST MLP and GAN implementations.
Implemented a full GPT model in dlgrad and successfully trained it on the TinyStories dataset to generate text.
Supports saving and loading models in safetensors format. Pip-installable.

GitHub →

gradtop

Real-time Neural Network TUI

Built a terminal UI in Rust to visualise neural network training metrics in real time.
Wrote Python bindings to hook directly into deep learning training loops and stream live data.
Renders high-frequency updates — loss curves, gradients — directly in the terminal with minimal overhead.

fcount

High-Performance File System Counter

A high-performance command-line tool in Rust designed to outperform standard Linux utilities (ls, wc) for file system statistics.
Achieved a 3.4× speedup over GNU coreutils (ls) and a 1.7× speedup over Go-based alternatives (gdu).
Main optimisation: bypass redundant stat() syscalls by leveraging Linux d_type directly from the kernel buffer, reducing system/kernel CPU time by 50%.
CI/CD pipeline via GitHub Actions builds, tests, and releases cross-platform binaries for Linux (x86_64) and macOS (Apple Silicon).

Finding Exoplanets using Object Detection

Bachelor’s Final-Semester Group Project

Helped astronomers analyse large volumes of light curve data from the TESS satellite.
Used the YOLOv5 algorithm to detect dips in light curve images.
Final model achieved 82% mAP and accurately detected dips on the test set.
Published as first author in Astrophysics & Space Science, Springer 2023 — see Research below.

Modelling Superconducting Flux Pumps for Fusion

MSc Dissertation — University of Glasgow

Investigated whether machine learning models can replace numerical models for simulating flux pumps in fusion applications.
Numerical models are computationally expensive, making them impractical for heavy fusion simulation workloads.
Trained 5 ML models on a dataset generated from numerical simulations.
Concluded with strong evidence that ML models are a viable alternative to traditional numerical methods for this domain.

Let’s talk.