Spotlight on Graduate Students: Suvaditya Mukherjee

Published: June 4, 2025
Category: News | Essays
Suvaditya Mukherjee

ICT is running a series of articles to highlight the work of our Graduate Research Assistants. In this essay we hear from Suvaditya Mukherjee, Masters student in Computer Science (Artificial Intelligence) at the USC Viterbi School of Engineering, while working as a researcher in the Learning Sciences Lab under the supervision of Dr. Benjamin Nye.

BYLINE: Suvaditya Mukherjee, Masters Student, Computer Science, USC Viterbi School  of Engineering; Researcher, ICT’s Learning Sciences Lab

From Sci-Fi Dreams to System Design

I first encountered the term “artificial intelligence” not in a classroom or a lab, but buried in the pages of a sci-fi novel. I was ten years old, reading about autonomous spacecraft, sentient assistants, and cities that thought for themselves. It was 2012, and the book optimistically predicted AI would be real by 2040 or 2050—futuristic, distant, and implausibly shiny. What it didn’t anticipate was just how soon fiction would blur into reality.

Fast forward a decade, and I found myself majoring in Artificial Intelligence at NMIMS University in Mumbai. Choosing the discipline wasn’t a carefully weighed decision—I saw “AI” on the list of options and instinctively selected it. That youthful spontaneity, however, was soon overtaken by the complexity and depth of what I had signed up for. Far from merely building robots, I was learning to build understanding—systems that could reason, learn, generate, and adapt.

An AI Awakening in the Classroom

The shift from curiosity to deep engagement happened in early 2023, when a few faculty members at my university began experimenting with ChatGPT in their classes. It wasn’t widespread yet, but the potential was immediately obvious. The professors were probing it with questions from our curriculum—asking it to simplify abstract algorithms or generate summaries of difficult concepts. I watched, fascinated, as an LLM turned into an academic aide, compressing complex theories into conversational clarity.

That moment ignited something enduring: an interest in how AI could serve as an educational partner. Education is often considered the great equaliser, but it’s only as effective as its accessibility. What if AI could lower the barriers—by translating, simplifying, personalising? I began to imagine models not just answering questions, but designing the questions themselves.

Crossing Continents: From Mumbai to Los Angeles

By the time I graduated, I knew I wanted to pursue graduate studies where AI was not just a subject of research but a practical tool for social impact. The University of Southern California became my destination, and in August 2024, I landed in Los Angeles to begin my Master’s in Computer Science (Artificial Intelligence).

Arriving here, I immediately reached out to professors whose research aligned with the vision I had begun forming. I was particularly drawn to translational work—projects that moved beyond benchmarks to address real-world applications, particularly in education, creativity, and generative AI. Professor Benjamin Nye at USC’s Institute for Creative Technologies (ICT) was generous in responding. We had a promising discussion, and soon after, I joined the Learning Sciences Lab as a student researcher.

Designing AI that Designs Curriculum

At ICT, I’m working on multi-turn and staged-generation systems using closed-source large language models such as OpenAI’s ChatGPT and Google’s Gemini. Our goal: develop automated workflows for generating diverse and pedagogically effective course content.

Here’s how it works. We begin with a curriculum artifact—a PDF chapter or instructor-generated notes. From that input, the system generates a range of educational components: question-answer pairs, multiple-choice assessments, hint-based scaffolding, and even sub-question trees. These aren’t static templates; they’re tailored, multi-layered structures that can adapt to different learning paths.

We’re especially interested in how LLMs can be coaxed into better instructional design through thoughtful prompting, modular pipelines, and iterative validation. LangChain plays a critical role in building these generative chains, allowing us to craft educational agents that don’t just inform—they teach.

This work feeds directly into the AIRCOEE initiative—a collaborative effort with the U.S. Department of Defense to democratise AI education for emerging technical personnel. We’re developing OpenTutor courses that not only explain concepts but simulate dialogue-based tutoring, making complex ideas more digestible for learners with diverse backgrounds.

Where Creativity Meets Code

Parallel to my work at ICT, I also hold a part-time position with USC’s School of Cinematic Arts, working under Professor Mark Bolas in the Interactive Games Division. Here, I’m focused on developing an introductory Python course for game developers, but the real fun begins with the research: applying generative models—LLMs and diffusion models—to automate creative workflows in media.

This dual engagement—working at the intersection of education and creativity—has allowed me to test how AI can operate across different epistemologies. Whether designing a calculus quiz or breaking down the narrative arc of a short film, the underlying question remains the same: how can machines help humans learn, imagine, and make meaning?

Scaling the Research Ladder

Before USC, my research trajectory included internships that shaped my technical versatility. At HARMAN International, I worked on a K-shot, rotation-invariant object detection pipeline, increasing accuracy by 35% on real-world client datasets. I also explored zero-shot time-series forecasting using LLMs and built an agentic system on Gemini 1.5 Pro to reduce cloud costs through intelligent spot-instance allocation.

At IIIT-Hyderabad’s Centre for Visual Information Technology, I tackled domain adaptation in autonomous driving, working with Professors C.V. Jawahar and Shankar Gangisetty. My focus there was on semantic segmentation and data transfer across domains—problems that are crucial as we move towards generalisable AI.

Earlier still, I interned with UnifyAI (formerly Ivy) in the UK, where I built demos for model conversion pipelines and managed documentation as a Google Summer of Code organization admin. One of the more experimental projects involved prototyping an LLM-based AI developer that could iterate on and expand codebases autonomously.

Each of these experiences taught me something different—how to write resilient code, how to handle scale, and most importantly, how to balance innovation with impact.

Speaking, Sharing, Scaling

As a Google Developer Expert in Machine Learning, I’ve had the opportunity to travel, speak, and collaborate across various technical communities. The most memorable experience to date was presenting at the 2024 PyTorch Conference in San Francisco.

In joint work with Shireen Chand from USC’s Information Sciences Institute, we demonstrated performance optimisations for 3D generation models using PyTorch. The talk focused on mesh-based generation and how native optimisation techniques could yield significant speed-ups without compromising quality. Presenting to an audience of deep learning practitioners, sharing something we built, and seeing it resonate—that remains one of my proudest moments.

Although I haven’t presented my ICT work publicly yet, I’m actively preparing for upcoming venues. NeurIPS, especially, is on my radar. This year it’s in San Diego—close enough to hope. I’d love to attend, especially if I’m able to present something meaningful. Fingers crossed for a chance to attend it!

Life Between the Labs

While my academic and professional life is intensely focused, I’ve found small joys in being part of the ICT community. Ice cream Thursdays are a personal favourite—not just for the sugar rush but the chance to exchange ideas casually (and regularly catch up with Ben!). Then there’s the time I took a Waymo One (self-driving car) to campus before its public rollout. As one of the early whitelist testers in LA, I rode from University Park to Playa Vista in a fully autonomous vehicle, watching heads turn as people noticed there was no driver.

Another cherished memory is when my parents visited from Mumbai in September 2024. It was only my second month in Los Angeles, and I brought them to the ICT building. They loved the campus atmosphere, and I introduced them to HomeState for their first Tex-Mex breakfast taco. They’re visiting again in June, and this time I have a shortlist of taco places ready.

Looking Ahead: Research and Resolve

As I head into the second year of my Master’s program, the inevitable question looms: what comes next? I’m exploring PhD opportunities, both within USC and elsewhere. ICT, of course, remains high on my list—few places offer the same blend of technical depth, application focus, and community.

I’m also looking into industrial research labs, especially those that champion open science, creative tools, or AI for learning. Whether in academia or industry, I want to keep building systems that not only work, but help others work better—whether they’re students, scientists, or artists.

The future I read about in that sci-fi novel has arrived, though perhaps not in the form I imagined. We may not yet have flying cars or interplanetary cities, but we do have machines that can learn, translate, and even teach. And as I continue my journey, I hope to help shape these systems into tools of empowerment—bridges between knowledge and those who seek it.

Publications and Research

• Presentation: Pushing the Performance Envelope : An Optimization Study for 3D Generative Modelling with PyTorch: Work on finding techniques to optimize 3D Text-to-Image Mesh generation [Accepted at PyTorch Conference 2024]

• Paper: Guiding the Student’s Learning Curve: Augmenting Knowledge Distillation with Insights from GradCAM: Work on investigating the effects of using GradCAM representations of Teacher models as direct inputs to Student models for quicker convergence. [Accepted]
• Paper: Project Lingua Franca: Democratizing Information through Unified Optical Character Recognition and Neural Machine Translation: Work on combined Optical Character Recognition and Neural Machine Translation for information translation with high-impact languages as targets [Accepted]

//