M.S. Computer Science · UMass Amherst

Hi! I'm Charan, a Machine Learning Engineer and researcher.

I'm doing my master's in CS at UMass Amherst. Lately, I've been talking to AI agents more than I talk to humans :/ I like knowing how things work all the way down, from the math to the metal. An interesting problem with a few rabbit holes and something to build? I'm probably in.

Looking for Summer 2027 internships and Spring 2027 co-ops in ML.

Contact LinkedIn Resume (PDF)

Show circles

From math to metal

I rarely stay inside one box

My degree was electronics. My minor was AI. Most of my work sits where the layers meet. A model that has to fit on a chip, a signal that has to stay clean, a pipeline that has to ship.

Machine learning Weakly supervised pathology models, medical image segmentation, model compression DeepMSI ADMM pruning
Software Data pipelines and tool workflows for agentic AI, containerized deployment, C++ build systems Comcast ARCO
Hardware FPGA prototyping, IMU motion tracking, models sized for edge devices ADMM pruning Medsby
Signals Band-limited oscillators, real-time flow features from network switches ARCO P4 / SDN

Selected work

What I built, what broke, what I learned

FeaturedFirst author · Accepted at ICAIHC 2026Jul 2025 – Jan 2026

DeepMSI

Predicting microsatellite instability in endometrial cancer from whole-slide histology

The problem

Microsatellite instability (MSI) is a tumor marker. A whole-slide image is gigapixels in size, and the only label is one per slide, MSI or not. There are no pixel annotations to learn from.

What I built

A weakly supervised pipeline. Each slide is cut into 256×256 patches, a ResNet50 turns each patch into a 2048-number feature, and attention-based multiple instance learning (MIL) learns which patches matter. A graph convolutional network (GCN) over the 8 nearest neighbours of each patch adds the spatial context a bag of patches loses. The attention weights double as maps a clinician can inspect.

What I learned

Working with patient data changed my priorities. Accuracy mattered, but so did interpretability and how the model fits into a clinician's workflow.

Pipeline, as described in the paper. My own diagram, with no patient images.

AUROC: 97.13%
Accuracy: 92.66%
Recall: 96.67%
F1: 87.88%

Validation set. Dataset of 542 whole-slide images.

First author · Manuscript in preparationJun 2025 – Jan 2026

ADMM structured pruning of nnU-Net

Shrinking a breast ultrasound segmentation model for memory-limited devices

The problem

nnU-Net segments breast ultrasound well, but at 46.3M parameters and 353.65 MB it is far too heavy for point-of-care or embedded hardware.

What I built

A three-phase pipeline. ADMM optimisation pushes whole filters toward zero over three rounds while the penalty tightens. Structural surgery then physically removes the dead channels from every layer and rewrites the network's plan. Fine-tuning for 50 epochs lets the survivors recover. I prototyped it on a small PyTorch CNN (MNIST, 80% sparsity) before scaling up.

What broke

I wrote a rule to protect the early encoder layers, which carry fine texture. It never fired. My layer-name pattern didn't match nnU-Net's real parameter names, so every layer was pruned at the same rate. The model held up anyway, but I now check that a safeguard actually runs. Also, removing channels barely sped up the GPU, because the network still does the same number of spatial operations. The win is memory, which is what an FPGA or edge device needs.

Parameters46.3M

after2.27M

Model size353.65 MB

after8.76 MB

Bars to scale.

Parameters: −95.1%
Model size: −97.5%
Dice: 0.766 → 0.752
HD95: −31.5%

BUSI dataset, nnU-Net 2D. Averages over 30 runs on the validation split. HD95 in pixels.

Comcast · AI InternMar – Jul 2026

Agentic AI remediation platform

Built data pipelines and tool-based workflows for the platform using ADK and Langfuse.
Migrated and containerized it with Docker and Kubernetes, and documented its architecture and setup.
Evaluated open-source frameworks for AI-driven virtual product placement and video editing.

Carnegie Mellon · Research InternFeb – Dec 2025

Band-limited oscillators for ARCO

Implemented band-limited oscillators in C++ for ARCO, a real-time sound synthesis engine, under Dr. Roger B. Dannenberg.
Used polynomial approximation and lookup tables to reduce aliasing.
Fixed CMake build issues to bring the engine to Windows.

Aliasing is what happens when a waveform carries more harmonics than the sample rate can hold. The hello at the top of this page is the same idea in reverse, a sum of harmonics with one circle each.

ProjectJul 2025 – Jan 2026

P4-assisted DDoS detection in software-defined networks

Designed a P4 pipeline that extracts flow features in the data plane, in real time.
Trained Random Forest and CNN detectors for low-and-slow (Slowloris) attacks.

F1, Random Forest: 96%
F1, CNN: 91%
Detection: ~25 s

Experience

Where I've worked

Mar – Jul 2026
Comcast, AI Intern Chennai, India
Data pipelines, tool workflows and containerized deployment for an agentic AI remediation platform.
Feb – Dec 2025
Carnegie Mellon University, Research Intern Remote
Band-limited oscillators and Windows support for the ARCO sound synthesis engine.
May – Jun 2025
Medsby Healthcare and Engineering Solutions, Project Engineer Intern Coimbatore, India
IMU-based motion tracking with a real-time dashboard, and a Blender character rigged to the sensor data for patient rehabilitation.
Jul 2024 – Dec 2025
AIMA Laboratory, Amrita Vishwa Vidyapeetham, Research Assistant Coimbatore, India
Co-led de-identified ultrasound and mammogram data collection with AIMS, Kochi. Trained nnU-Net models for breast ultrasound tumor segmentation (Dice 0.766 on BUSI), scheduled for pilot testing at Cachar Cancer Hospital, Assam.

Education

School

University of Massachusetts Amherst
M.S. Computer Science · 2026 – 2028
Machine Learning, Neural Networks, Software Engineering
Amrita Vishwa Vidyapeetham
B.Tech. Electronics and Communication Engineering, minor in AI and ML · 2022 – 2026

Publications

Papers

DeepMSI: A Deep Learning Framework for Predicting Microsatellite Instability from Whole-Slide Histology and Tumor Genomic Markers for Endometrial Cancer
First author · Accepted at ICAIHC 2026
ADMM-Based Model Compression for Real-Time Breast Ultrasound Segmentation on FPGA Systems
First author · Manuscript in preparation

Tools

What I use

Python, C++, C, PyTorch, TensorFlow, scikit-learn, Hugging Face, OpenCV, nnU-Net, LangChain, ADK, Langfuse, Docker, Kubernetes, CUDA, Linux, Git, CMake, P4, VHDL, FPGA (Basys3 Artix-7)

About

A bit more about me

I started out in electronics because I wanted to know how a signal gets from the real world into a computer. That question never really stopped. It pulled me into AI, then into the limits of hardware, and now into a master's in computer science.

A lot of what I know came from things not working. A model too big for the memory it had to live in. A build that refused to compile on Windows. A safeguard in my own code that turned out to never run. Each one sent me down a rabbit hole, and I usually came back with something better.

I like work that stays close to the people who rely on it. That's why so much of mine has been in healthcare, where a model only helps if a clinician can trust what it shows.

Off the clock

Outside the lab

Tennis
Basketball
Badminton
Pickleball
Hiking
Photography
Gym rat

Still on my list to try are golf, skiing and snowboarding.

I want to hike more mountains. Next on the list is Mt. Washington, so if you're going, I'd happily come along :)

Contact

Open to Summer 2027 internships and Spring 2027 co-ops in ML

Available for co-ops from January to August 2027.

cuthaman@umass.edu LinkedIn Resume (PDF)