Teaching

This course is about what happens after the model works in a notebook. Real memory limits. Real power budgets. Real latency constraints. Students take ML ideas and figure out how to make them run on hardware that doesn't have the luxury of a GPU.

I worked alongside Banhimitra Kundu and Ketan Chaudhary, helping students get TinyML inference running on microcontrollers. That meant dealing with quantisation trade-offs, debugging real-time pipelines, and spending a lot of time figuring out why something that worked perfectly in simulation fell apart on actual silicon. The course is built around learning by building, and the final project showcases were consistently the highlight.

This is a course designed for people who want to understand the mathematics behind machine learning properly, not just memorise formulas and move on. It covers linear algebra, calculus, optimisation, and probability, the pieces that hold everything together once you go past the surface level of ML.

My contribution was primarily visual. A lot of mathematical ideas are easier to understand when you can actually watch them unfold, so I built animated visualisations using Manim to bring the concepts to life. Things like how eigenvectors behave under transformations, what SVD looks like geometrically, how gradient descent navigates a loss surface, how projections work in high-dimensional spaces. The kind of intuition that a static equation on a blackboard struggles to convey.

The goal was always to help learners build real conceptual understanding rather than fall into formula-based pattern matching. If you can see what a matrix transformation does to a shape, you don't need to memorise the rules. You just know.

I'd recommend this course to anyone beginning ML, engineers transitioning into data science, or really anyone who wants mathematical clarity rather than mathematical anxiety.

TalentSprint runs a deep learning program for working professionals and industry engineers. Within that program, there are a few dedicated sessions on TinyML, and that's where I come in. My role was to help participants understand what changes when you move from training models on powerful machines to deploying them on a microcontroller with 256KB of RAM.

Together with Banhimitra Kundu and under Prof. Thakur's guidance, we ran these sessions across five cohorts, entirely online. Sixty people in a call, each with a different hardware setup, trying to flash firmware and debug sensor pipelines remotely. It was chaotic in the best possible way. And it taught me more about patience, clear communication, and designing learning experiences that actually work than any classroom setting ever could.

At Vidyakosa, the audience is different. These are undergraduates still building their foundations. I help teach microcontroller programming, sensor integration, communication protocols, and deploying ML models on constrained devices. But the deeper theme running through all of it is compute and power awareness. Understanding why efficiency matters. Why engineering choices have real consequences. Why "it works on my laptop" is not the same as "it works."

These sessions are exploratory. Students build small systems, experiment, break things, and debug together. Watching young engineers grow confident with hardware and discover that building things is genuinely fun never gets old.

This is where a lot of my thinking about education took shape. CFAL runs learning environments for younger students. Robotics, early coding, electronics projects, open-ended tinkering. I started as a Makerspace Coordinator and later moved into an Academic Coordinator role.

The work wasn't technically complex. But it was formative. It taught me about pacing, about meeting learners where they are, and about creating spaces where students feel free to explore rather than perform. More than any other experience, CFAL shaped how I think about what it means to teach well.