Rakshith Mahishi

I was a Teaching Assistant for the course EE260C: Introduction to Self-Driving Stack teaching page where I was responsible for releasing labs on perception and planning, conducting office hours and grading every students labs and projects. I also conducted lectures from time to time.

I was also a reviewer for AAAI 2025 where I reviewed for AAAI 2025 workshop on the topic Scalable and Efficient AI Systems.

I was a grader for Foundations of Machine Learning (CS224) where I graded assignments and exams of a class of 70.

I was a part of the program committee for ICCV workshop: Anomaly Detection with Foundation Models (ADFM) Official Page

• Self-Driving Stack: i. Implementing an out-of-distribution detection model by using energy score as a metric to an object detection dataset in the self-driving Domain.
  ii. Implementing a new uncertainity metric as a method of evaluation for motion planning. This method once verified can be extended towards prediction, mapping and localization.
  [Current Research]


• Diffusion Models for Point Process: Currently developing diffusion models for point datasets to recreate certain paths in a high-dimensional space using an unconditional framework where the number of points are unknown.
  [Current Research]


• Jailbreaking Language Models: Developed several techniques to jailbreak the LLAMA2 (7B and 13B) model to answer harmful questions from the AdvBenchDataset. [Source Code] [Paper]


• Building GPT2 from scratch: Inspired by Andrej Karpathy's course on building GPT2 124M model from scratch by implementing several components like PyTorch's autocast and GradScaler, and created unit tests using pytest. [Source Code]

• MASTERS THESIS: Diffusion Models for Point Processes(2025)
  
  collaborators: Rakshith Mahishi, Prof. Christian R Shelton.
  
  Abstract: We propose a novel diffusion-based framework for modeling and denoising spatial point process data, with a specific focus on urban mobility patterns. Unlike traditional diffusion models that assume fixed cardinality and treat individual points independently Ho et al. [2020], our method represents entire point sets as unified entities, enabling the learning of spatial dependencies and variable cardinalities.
  [Report]


• A SYSTEM AND METHOD FOR ENABLING A SMART CUBICLE(2021) Number - 202341012174 (Patent Approved)
  
  collaborators: Prof. Sahana D Gowda, Rakshith Mahishi, Praneeth Kumar, Nandana V, Param S, Prajwal Rajegowda.
  
  Abstract: Built a plug and play home automation system that is app controlled and any service such as switching on fan or turning on light can be controlled via phone. The novelity lies in the plug and play design where users can just plug in a necessary component and enable the specific service.
  [View Status][Report]


• A Survey Paper on Cancer Cell Classification using Deep Learning.
  
  collaborators: Rakshith Mahishi, Praneeth Kumar, Nidhi Srivastava, Prof. Chayadevi M L.
  
  Abstract: This paper contains different methods employed by different authors for classifying cancer cells mainly breast cancer cells. We then provide our novel idea on how this problem can be tackled and achieved in a different way. The novelty of this lies in the dataset we manually collected from cancer hospitals and pre-processing them for training.
  [Paper]