Arnav Kulkarni
Software Engineer × Full-Stack Developer
अर्णव | Sanskrit: Ocean, Sea
I am a software developer and machine learning enthusiast with hands-on experience in research, systems engineering, on-device deep learning, and generative AI. I am excited to contribute to innovative projects that make a meaningful impact.
Work Experience
Staff Research Associate II
UCSD, Contijoch Research Laboratory
Engineered a C++ package for Siemens MRI scanner to implement Autonomous Radial K-space Sampling (ARKS), improving uniformity by 15% and image quality by 6% for cardiac MRI through data-driven angle selection.
Utilized nested virtualization to compile and validate shared libraries, enabling interoperability between the reconstruction system running on Linux and the Windows-based control environment.
Implemented a high-performance compute pipeline with efficient memory and buffer management, ensured scalability, while optimizing for low latency and high throughput in real-time imaging.
Research Assistant
UCSD, Machine Intelligence, Computing and Security Lab
Compressed model size without altering model architecture by collapsing layers of the neural network.
Implemented LayerCollapse Algorithm on Integer-only BERT to selectively collapse layers of the Transformer model.
Optimized model training with CUDA for GPU acceleration, and conducted debugging and performance evaluation.
Software Developer
UCSD, ITA Conference
Engineered dynamic event scheduling system with streamlining planning and engagement.
Designed a database for 6000+ attendees, implemented authentication based on OAuth2, access control, and integrated RESTful APIs to suppport web and mobile application.
Designed and developed a responsive front-end, improving user experience across devices.
Research Assistant
UCSD, Boolean Labs
Designed a lightweight neural network for on-device arrhythmia detection using time-series ECG signal, deploying on embedded hardware.
Deployed compressed models on STM32-microcontroller for real-time detection of Ventricular Arrhythmia in patients from ECG signal.
Implemented selective pruning based on Lottery Ticket Hypothesis, retaining only critical model weights, reducing memory usage by 37% and maintaining 97% precision.
Utilized Docker and Kubernetes to manage distributed training across GPU nodes.
Research Assistant
PICT, Computational Linguistics Lab
Conducted research on multi-modal video summarization of English news videos to develop an assistive tool for the visually impaired.
Utilized seq2seq RNNs for subtitle generation
Used CRAFT (Character Region Awareness for Text Detection) for text detection in video frames with WER of 0.96.
Summarized the generated text using BERT-based transformer models.
Software Development Intern
Siemens Digital Industries Software
Improving user experience in finding items unassigned to structure partitions in CAD designs on the web-based client PLM platform, Teamcenter.
Developed a proof-of-concept feature that enhanced item finding functionality for unassigned items in CAD designs.
Refactored feature test files to improve reliability of Cucumber tests for Acceptance Test-Driven Development (ATDD).
Software Development Intern
Dynamisch IT Pvt. Ltd.
Developed a client-centric dashboard for visualizing real-time financial data, and enhanced the web application's performance with regard to the response time.
Developed a dynamic dashboard and integrated RESTful APIs to ensure seamless updates.
Optimized API calls to improve data load times, enhancing user experience and dashboard responsiveness.
Featured Projects
A showcase of my recent work and side projects
Built a cloud-based file storage service with distinct microservices for efficient data storage and metadata management. Implemented the RAFT consensus protocol across metadata servers to improve fault tolerance and ensure high system uptime and consistency.
Designed and implemented a HTTP/1.1 web server using socket programming. Developed multi-client capability with persistent connections, request pipelining, and timeout mechanism to optimize network efficiency. Ensured secure file access with URL validation.
Developed a Question Answering model using Llama, optimized through prompt engineering and chain-of-thought reasoning. Utilized Parameter Efficient Fine-Tuning (PEFT) techniques like LoRA to optimize generative capabilities. Integrated a RAG module using FAISS and LangChain for better contextual understanding.
Hey Ambulance!
Devised an intelligent traffic control system using Atmega-328 microcontroller capable of operating stoplights to create a dynamically shifting green corridor for ambulances based on GPS location. Built a mobile application to call ambulances and manage patient data on Google Cloud. Selected for finals (top 10) at Rakuten Hackathon from 6000+ teams.
Skills & Expertise
Technologies and tools I use to bring ideas to life
Languages
Frameworks & Libraries
AI/ML
Tools & Cloud
Let's Connect
I'm always open to discussing new projects, creative ideas, or opportunities to be part of your visions. Let's build something amazing together!