Building machine learning evaluation pipelines, deep learning experiments, and practical AI systems.

Training Dynamics

Analyze how neural networks learn over time across optimizers, architectures, and data settings.

Generalization and Stability

Design metrics and experiments to study robustness, convergence behavior, and reliable performance.

Model Analysis and Evaluation

Build tooling to inspect eigenspectra, alignment, internal module behavior, and training bottlenecks.

Research Engineering

Develop reproducible PyTorch workflows for model evaluation, debugging, and large-scale experimentation.

Languages

Python, C, SQL

ML & Research Frameworks

PyTorch, Functorch, OpenCV, Hugging Face Transformers, NumPy, Pandas, Scikit-learn

Research Areas

Deep Learning, Model Evaluation, Training Dynamics, Reinforcement Learning, Multimodal Learning, Quantum ML

Tools

Git, Linux, MySQL, Qiskit, Gymnasium, Stable-Baselines3, ELK Stack, Kibana

NTK Alignment and Training Dynamics Across Modern Architectures

Built PyTorch-based experimental pipelines to analyze optimization and generalization behavior across MLPs, CNNs, vision transformers, and GPT-style models. Studied eigenspectra, label-kernel alignment, and convergence trends to compare theoretical predictions with empirical training behavior.

Module-Level Transformer Analysis Pipeline

Developed custom Functorch-based tooling to inspect attention and MLP blocks at the module level, track internal training behavior, and identify bottlenecks affecting optimization. Used these analyses to support architecture-level comparison and experiment design.

Transformer-Based Text-to-Image Generation

Developed a text-to-image generation model inspired by CogView using Python and PyTorch. Built a custom VQ-VAE autoencoder and transformer model trained on MS COCO, incorporating PB Relax and Sandwich Layer Normalization to stabilize training.

External Memory for Reinforcement Learning

Developed a notebook-based external memory structure for reinforcement learning using Python, Gymnasium, and Stable-Baselines3. Improved rewards by 14.5% and reduced task completion time by 65.5% in partially observable environments.

Quantum Diffusion Denoising Probabilistic Model

Re-implemented QuDDPM using Qiskit and Python, leveraging low-depth quantum circuits for generative modeling and evaluating output quality using Maximum Mean Discrepancy loss.

AI WhatsApp Bot with FastAPI and Twilio

Built a WhatsApp chatbot prototype using FastAPI, Twilio WhatsApp Sandbox, and lightweight local language model integration. Designed the webhook-based backend, request handling flow, and response generation pipeline for conversational automation.

Purdue University · Graduate Researcher

Aug 2023 – Present

Led PyTorch-based experiments analyzing training dynamics and generalization across vision transformers, CNNs, and GPT-style models. Built custom Functorch pipelines to track module-level behavior, diagnose bottlenecks, and guide optimizer and hyperparameter choices. Designed evaluation metrics to study training stability and inductive bias for more reliable cross-architecture comparisons.

Purdue University · Graduate Teaching Assistant

Jan 2025 – May 2026

Supported lab instruction, grading, and student guidance across coursework in graphical communication, spatial analysis, and power electronics. Helped students debug technical work and improve engineering communication.

Zoom Video Communications · Service Engineer

Jul 2022 – Jul 2023

Analyzed and resolved real-time production ML issues across Zoom product pipelines using ELK Stack, Kibana, and MySQL. Collaborated with research, development, and operations teams to design internal automation tools using Python and C, reducing issue resolution time by 65% and improving support efficiency for premium enterprise clients.

UST Global · Data Science Intern

Jun 2021 – Jul 2021

Designed and developed an OpenCV- and machine learning-based interface to automatically generate electricity meter bills from uploaded meter images. Achieved 100% automation and reduced manual processing time by 70%.

WiE 5th Annual Symposium, Purdue

Presenter · February 2026

Presented research on training dynamics and label-NTK alignment in modern neural networks.

arXiv Preprint

May 2026 · arXiv:2605.25275

Label-NTK Alignments and A Tighter Convergence Bound in the NTK Regime
Ruchirinkil Marreddy and Chaoyue Liu

We develop a tighter convergence analysis in the NTK regime and study how label- and residual-NTK alignments yield more informative predictions of neural network training dynamics than worst-case spectral bounds alone.

I like connecting theory and practice

I enjoy turning mathematical questions in deep learning into experiments, plots, and testable engineering insights.

I work across multiple AI areas

My experience spans deep learning theory, reinforcement learning, multimodal generation, quantum ML, and practical AI tooling.

I enjoy building usable systems

Beyond research, I like building end-to-end prototypes that turn technical ideas into real workflows and applications.

I care about communication

I enjoy presenting technical ideas clearly and making complex work understandable for different audiences.