Dhawal Modi

I am an MS student in Electrical Engineering & Computer Science graduate group at University of California - Merced.

Before joining UC Merced, I worked at Tata Consultancy Services as a Backend Software Developer, where I developed payment processing pipelines and integrated them with third-party payment settlement systems (HKICL, BPay to name a few).

My technical skills include Python, Java, C++, and frameworks such as PyTorch, TensorFlow, and Spring Boot. I am also proficient in using technologies like ROS, Django, Flask, Oracle SQL, IBM Db2, CUDA. My experience spans from designing microservices architectures, integrating REST APIs, to fine-tuning LLMs and Deep Learning models.

I have a Bachelor's in Engineering in Electronics & Communications Engineering from Rajiv Gandhi Proudyogiki Vishwavidyala.

Email  /  CV  /  LinkedIn  /  Github

My academic focus and hands-on experience are in machine learning, deep learning, and computer vision. My current research involves developing and implementing person-following and tracking capabilities for real-time robotic applications. Additionally, I also dabble in optimizing Deep Learning model performance on embedded platforms like the Nvidia Orin AGX.

• Developed a mobile robotic platform using the AgileX SCOUT UGV to assist forest crews in reducing wood waste and preventing wildfires.
• Finetuned ENet and SegFormer Semantic Segmentation models for Robot path traversibility application on custom dataset with an mIOU of 0.729 and 0.68 respectively (improvement of 23% and 15% over baseline models).
• Trained and deployed ENet CNN model on Nvidia Orin AGX for real-time tasks, achieving 25ms inference time per frame.

• Implemented and streamlined validation routines for Outward Direct Debits and Credit Transfer payment channels, developing interfaces as per FPS and HKICL specifications. This enabled high-efficiency processing, handling up to 100,000 transactions per minute.
• Designed Spring REST API Client integration with BPAY API, enhancing real-time bill payment processing by 35%; led microservices architecture development for rapid API prototyping, ensuring successful deployment to testing and production environments.
• Conducted comprehensive design sessions with clients and users for Inward/Outward Direct Debit modules, resulting in precise documentation that enhanced developer accuracy by 30% and cut module testing duration by 5 hours a week.
• Led a cross-functional team of 3 SDETs in implementing a Spring Boot API testing pipeline across 3 testing regions, leading to a 20% increase in code delivery speed.
• Innovated and created an automated testing framework for 15 user stories and application flows using Jbehave BDD Framework, substantially improving regression testing efficiency.
• Implemented REST APIs in Java Springboot to serve payments processing logic for Credit Transfers, Direct Debits, ACH, and Mandates.
• Developed and deployed mock webservices for Temenos and T24 Bill payments API integrating TCS Bancs logic.

• Teaching assistant for CSE-015 Discrete Mathematics.
• Responsible for conducting lab sessions, grading quizzes, homeworks, assignments and exams along with providing academic assistance for a class of 90 undergraduate students.

Led a team of 3 undergraduate students to implement and design Machine Learning and Deep Learning models for Heartbeat classification and Transmembrane Potential curve reconstruction.

Implemented Customized 1D SqueezeNet CNN model to predict Myocardium Activation Times and Transmembrane Potential Curve reconstruction with 97.46% accuracy on 1600 samples of unseen data.

The goal of this project was to predict a prompt, ⟨Rewrite Prompt⟩, that was used to generate ⟨Rewritten Text⟩ by providing only ⟨Original Text⟩-⟨Rewritten Text⟩ pairs to the LLM.

During inference, it predicts an embedding using the ⟨Original Text⟩ and the ⟨Rewritten Text⟩. This embedding is compared against a knowledge database of embeddings. The most similar embedding in this database is our model prediction. The score for each predicted / expected pair is calculated using the Sharpened Cosine Similarity, using an exponent of 3.

Our finetuned models ranked 986 out of 2186. The best leaderboard score was our Phi-2 submission which scored 0.6188. The Mixtral 8x7B model finetuned with 1000 samples scored 0.6056 and Gemma 7B finetuned with 5000 samples scored 0.5981.

The goal of this competition is to use various factors to predict obesity risk in individuals, which is related to cardiovascular disease.

For this Episode of the Series, the task is to predict whether a customer continues with their account or closes it (e.g., churns).

The Problem : As cities grow, traffic gets worse. Poorly planned infrastructure and improper traffic policies lead to traffic bottlenecks. Traffic jams and long wait times at intersections lead to wastage of time and fuel and can cost a country in terms of it's GDP.

The Solution: Real-Time Dynamic Traffic Management system. Using Vehicle Detection and Lane occupancy we can adjust to real-time traffic scenarios. The system detects traffic conditions using vehicle count ,vehicular density and delay time using OpenCV. Traffic light timing is then propogated to Arduino controlled traffic lights that can adapt to traffic conditions on a cycle-to-cycle basis.

More projects...

Website template credits go to Jon Barron.