Hi, I am Rishit Kar, a computer engineering undergraduate at DJ Sanghvi College of Engineering, Mumbai University.
Incoming Machine Learning Engineer Intern at IDfy, Asia's leading TrustStack.
During a summer visiting research internship at IIT Mandi, I explored machine learning for 3D propellers and geometry, including a literature review on ML for propeller design. I enjoyed the engineering challenge, but realized 3D ML wasn’t quite my direction 🙂. Photos 📷
Later, remotely at IIT Patna with Ms. Shivani Gupta, I worked on disease phenotyping with temporal and spatial GNNs, combining chest X-rays, clinical text, and EHR data from MIMIC. That experience helped me discover how much I enjoy machine learning and core deep learning.
In collaboration with Dr. Pritha Gupta at Ruhr University Bochum, we are writing a paper on how the order of LoRA fine-tuning affects factual recall in large language models. We test whether updates to a selected block of layers can preserve newly learned facts on their own, compared with the fully fine-tuned model. The paper is under review.
In collaboration with DJS ACM, we have published work in workshops at ICLR and ACML. I’m especially interested in modern agentic AI and LLM systems.
I was selected for the fully funded Google DeepMind APAC Research Symposium 2026 in Bengaluru (29–30 October). The sessions and conversations made it a memorable couple of days. Photos 📷
I enjoy competitive programming and have solved more than 400 problems across Codeforces and LeetCode. I would love to work in software engineering or machine learning roles, building useful systems and products for production.
Email  /  CV  /  Google Scholar  /  Github  /  LinkedIn  /  LeetCode
I am most interested in core deep learning, large language models, and agentic AI , their architectures, their theoretical foundations, and their practical applications across vision, language, and reasoning.
Under Review: EACL SRW 2027
Work in Progress: Mitigating hallucination and confabulation in multilingual settings
Neuro-inspired gated memory with utility filtering, episodic buffering, and pruned knowledge-graph consolidation. Evaluated on 100 HotPotQA samples across Qwen, Gemma, and Mistral, achieving F1 0.312 vs 0.068 baseline.
We benchmark frozen Reverso and Chronos-Bolt representations on six MONSTER datasets with 1%, 10%, and 100% of training labels. Pretraining beats random-weight controls on five datasets, but frozen features rarely match strong task-specific classifiers. Restoring amplitude information recovers some of the gap, while fine-tuning and forecast-error features do not close it.
Multi-modal cancer classifier on the MLOmics benchmark (8,000+ patients, 32 cancer types). Dynamic context gating achieved best precision: GS-BRCA 0.890, GS-COAD 0.947, GS-LGG 0.985.
Presented in front of 75 people including students and professionals. An introductory workshop on PySpark, distributed data processing, and practical big-data workflows.
A talk on finding research opportunities, writing effective outreach, and building a student research profile.
I geek out over FC Barcelona. From Messi to Lamine Yamal, I’m a true culé. Visca el Barça! 💙❤️
I used to love painting; it’s still one of the hobbies I remember fondly.