Aditya Roy Chowdhury

Education

Academic background.

Aug 2024 to May 2026

University of Southern California

Master of Science in Computer Science

GPA: 3.83/4.00

Coursework

Deep Learning and Its Applications
Analysis of Algorithms
Advanced Computer Vision
Applied Natural Language Processing
Security Systems
Machine Learning for Data Science
Information Retrieval and Search Engines
Software Engineering

Los Angeles, CaliforniaUSA

Aug 2020 to June 2024

Thapar Institute of Engineering Technology

Bachelor of Engineering in Computer Science and Engineering

GPA: 9.33/10.00

Coursework

Mathematics I
Mathematics II
Computer Programming
Object Oriented Programming
Data Structures
Operating Systems
Computer Architecture and Organization
Discrete Mathematical Structures
Numerical Analysis
Database Management System
Computer Networks
Design and Analysis of Algorithms
Software Engineering
Artificial Intelligence
Machine Learning
Optimization Techniques
Probability and Statistics
Network Programming
Cloud Computing
Edge AI and Robotics: Accelerated Data Science
Edge AI and Robotics: Data Centre Vision
Edge AI and Robotics: Embedded Vision
Edge AI and Robotics: Reinforcement Learning & Conversational AI
Computer Graphics
Quantum Computing
Theory of Computation
Cognitive Computing
Compiler Construction
Introduction to Corporate Finance

Patiala, PunjabIndia

Experience

Where I've built and shipped.

Feb 2024 to Apr 2024

International Centre for Theoretical Physics (ICTP)

Machine Learning Engineer Intern

Built a small AI model that detects objects in poor visibility and optimized it for edge deployment.

Developed a Keras/TensorFlow object detection pipeline for adverse visibility conditions, utilizing knowledge distillation to compress SpineNet 49S backbone by 50% by maximizing SSIM.
Enabled FPGA deployment by porting Python model inference logic to C++/HLS with Vivado, generating hardware-synthesizable code.
Collaborated with stakeholders to produce 10K LOC of Python/C++, building models and reusable tooling to streamline data processing and model evaluation.

TriesteItaly

Jun 2023 to Jul 2023

Thapar Institute of Engineering Technology

Machine Learning Research Intern

Led a small team building an app that monitors heart health in real time and alerts when something may need a doctor’s attention, using AI to spot irregular heart rhythms.

Led a 3-person engineering team through Agile sprints to build a full-stack health monitoring system, coordinating integration of a Flask backend with a responsive HTML/CSS frontend for real-time visualization.
Mitigated severe class imbalance in arrhythmia classification by constructing a tailored signal preprocessing pipeline, attaining 0.7 recall with a 1D ResNet.
Ensured instant notification for critical cardiac events while minimizing false positives by engineering a hybrid alerting mechanism that combines ML predictions with static vital thresholds.

Patiala, PunjabIndia

Jun 2022 to Jul 2022

INDUS TOWERS

Machine Learning and RPA Intern

Used data analysis and process review to find bottlenecks and scope automation, and supported an NLP project through technical reviews.

Identified operational bottlenecks by analyzing 6 months of historical ITSM data across 10,000+ support tickets using advanced pivot tables, resulting in accurately mapped volume patterns and quantified system downtime.
Defined the scope for an outsourced email triage automation system by evaluating historical support workflows, resulting in a targeted plan to automate 35% of recurring Level-1 requests.
Supported a senior manager in overseeing a vendor-built NLP project across 2 technical review cycles, formulating technical queries for stakeholder meetings to ensure proposed solutions addressed internal constraints.

GurugramIndia

Projects

Selected work.

Findly, file explorer and search interface

Findly: Privacy-First RAG Agent

A search tool that finds information across your documents and images while keeping everything on your own device, with no cloud and no data leaving your computer.

Architected a RAG pipeline with a dual-indexing backend to separate text (384d) and image (512d) vector spaces using FAISS and SQLite, achieving a 4.33ms mean search latency over 10K queries for CPU-only inference.
Constructed a synchronization pipeline using SHA256 file signatures and thread-safe locking to ensure index integrity, maintaining a 42.27ms/file indexing throughput.
Verified a memory-efficient core and currently undergoing integration of Sentence Transformers and CLIP to replace mock embeddings.

Link

CodeMaya, 3D render example (snowman figure)

CodeMaya: CGI Assistant

Describe a scene in plain English and the system turns it into a 3D scene you can render, like having an assistant that writes the technical instructions for you.

Spearheaded development of two-stage training pipeline, cascading Supervised Fine-Tuning with a multimodal alignment phase to train LLM to convert natural language prompts into executable scripts for 3D Renders.
Specialized CodeLlama-7B for domain-specific syntax by LoRA fine-tuning on scraped documentation and code, resulting in 92% syntax validity.
Distilled semantic knowledge from CLIP to align generated 3D renders with text prompts by integrating a visual feedback loop based on Contrastive Loss, yielding 91% structural similarity.

Link

Forecasting Parkinson's, UPDRS 2 predicted vs actual

Forecasting Parkinson's Progression

A system that helps predict how Parkinson’s may progress over time using only routine check-up scores, so doctors can plan care without extra tests. Research published in a peer-reviewed book.

Modeled non-linear disease trajectories to predict Parkinson's progression by devising an ML pipeline on a longitudinal dataset of 237 subjects.
Demonstrated a cost-effective alternative to complex biomarker screening by delivering 0.57 MSE on purely non-invasive features (UPDRS scores).
Validated findings by publishing research paper in Springer Lecture Notes in Computer Science.

Link

Dimming the Danger prototype, Arduino, relay module, LED assembly, servo

Dimming The Danger: Collision Avoidance System

A headlight system that automatically dims only the part that would glare at oncoming drivers and uses AI to spot cars in the dark so it knows when to dim.

Prototyped an Adaptive Driving Beam (ADB) system on an Arduino microcontroller, programming a zone-based control logic to selectively deactivate LED sectors for oncoming traffic.
Adapted YOLOv8 for low-light conditions by optimizing model parameters on a custom-curated dataset of nighttime driving scenarios, achieving 0.89 recall for vehicle detection.

Link

SpineNet49S Keras

Rebuilt a well-known Google AI image model so researchers and developers can use and experiment with it easily; shared publicly for free.

Reproduced Google Research's SpineNet architecture in Keras, implementing the novel scale-permuted backbone with cross-scale connections and releasing model for open-source access on Hugging Face.

Link

Beyond academics

Hobbies, volunteering, or side interests.

Photography

Things I find pretty or interesting.

Volunteering

Community involvement.

Get in touch

Open to opportunities and conversations.