Harshvardhan Aditya

Deployed diffusion, background removal, video likelihood estimator, image likelihood estimator, and upscaler models on a server with 4 T4 GPUs. Developed asynchronous API endpoints using FastAPI, Celery, Redis, and RabbitMQ. Built three product pipelines for text-to-video, text-to-image, and image-to-image generation, increasing text-to-video generation speed by 400%.

Project- Whispered Tuning: Guarding Training Data Privacy amidst Fine-Tuning of Language Models through Differential Privacy. Analyzed vulnerabilities in Large Language Models across key sectors. Worked at the School of Cybersecurity and Privacy on private fine tuning of LLMs, enhancing model utility by 53%. Implemented differential privacy techniques using HuggingFace, Python, PyTorch, LoRA, PeFT, and Opacus.

Worked under Prof. Tanveer Ahmed on LLMs and their applications in the field of Law. Built multi-tool agent question-answering system with self-correcting query engines. Analysed large amounts of legal data using NLP to help lawyers make more informed decisions.

Worked on video and image segmentation models and their inference time optimizations for edge devices. Used quantization and multi-threading for reducing video segmentation time for multiple frames, reducing model size by 200%.

Student Teaching Assistant- Computational Thinking & Programming with Python and Java. Taught Python and Java in Lab sessions to first year students. Mentored seven batches of 30 students each.

Ranked 2nd out of 518 students in the Computer Science Department. Letters of Appreciation for ranking in the top 1% through all semesters.