Ritesh Bhalerao

S-KANformer: Enhancing Transformers for Symbolic Calculations in High Energy Physics

Optimizing Biomass Forecasting & Supply Chain : An Integrated Modelling Approach

SineKAN: A Flexible Machine Learning Model for Modelling L² Functions and Beyond (Co-author)

Optimizing Biomass Forecasting & Supply Chain : An Integrated Modelling Approach (Presenter)

Research Scientist

• Working towards making AI more interpretable, aligned and intelligent.
• Building novel, hard-to-hack, diverse RL environments for frontier AI research.

ML Alignment and Theory Scholar (Exploration Phase)

• Selected through a highly competitive application process based on a research project in mechanistic interpretability.
• Worked under the mentorship of Neel Nanda in the mechanistic interpretability stream.
• Worked on a project focusing on studying evaluation awareness in open-source LLMs.

Research Assistant (Google Summer of Code)

• Selected for consecutive years to pursue research in machine learning applications for high-energy physics.
• Developed advanced Transformer-based models augmented with the Kolmogorov–Arnold Network for symbolic computation of particle interactions.
• Acquired hands-on experience with large-scale training on High-Performance Computing clusters, including Perlmutter.
• Currently exploring reinforcement learning fine-tuning techniques to enhance model alignment in symbolic physics tasks.
• Project link

Research Intern

• Worked on developing a multi-agent system for automatic vulnerability detection in entire code bases.

Summer Research Fellow

• Developed a domain specialist chatbot by Retrieval Augmented Fine-Tuning of Llama 2 - Meta’s open-source LLM.
• Created an end-to-end domain-agnostic pipeline, including a custom Q&A data generation workflow.
• Repository link

Jr Technical Officer

• Development of council’s mobile application & website.
• Conducted technical workshops for 100+ students.

Vivekanand Education Society's Institute of Technology, Mumbai

CGPA : 8.47

B.A Talreja College, Badlapur

Percentage : 94.67

Kaggle Featured Competition – PII Detection from Educational Data

Achieved a Bronze Medal in the Kaggle Featured Competition on detecting personally identifiable information (PII) in student essays.

Shell.ai global Hackathon (Agricultural Waste Challenge 2023)

Our solution for Biomass supply chain optimization for state of Gujarat using ML and MILP was ranked 12th on global leaderboard

E-Yantra Innovation Challenge

EYIC is a national level competition . My team built an app called "Enabled" which is community platform for persons with disabilities. Pitch link

IIIT Sri City National Healthcare Hackathon - 2023

My team's solution - "Rhythm" which is a web and app based integrated platform in the domain of preventive cardio vascular self-care won 2nd prize at this hackathon.

Evaluation Awareness in Open Source Models

Developed an automated pipeline to generate environments eliciting evaluation awareness in LLMs. Benchmarked multiple black-box and white-box suppression techniques.

Does Your LLM Trust You?

Investigated whether LLMs form internal "Trustworthiness" attributes of users and whether these can bypass safety guardrails. Trained linear probes on Llama models to extract trust vectors from synthetic multi-turn conversations. Demonstrated that trust vectors are mechanistically distinct from compliance/refusal directions and successfully induce jailbreaking through a novel mechanism—making the model perceive users as trusted individuals rather than directly suppressing refusal.

S-KANformer: Enhancing Transformers for Symbolic Calculations in High Energy Physics

Developed and applied S-KANformer (Transformers infused with Kolmogorov-Arnold Networks Using Sinusoidal Activation Functions) for high-energy physics symbolic calculations, achieving state-of-the-art performance.

Fragile Inceptions

Investigated belief fragility in hybrid reasoning model by fine-tuning Qwen3-1.7B on counterfactual data and evaluating behavioral robustness. Applied mechanistic interpretability using BatchTopKCrossCoder, identifying fine-tune-specific latents linked to incepted counterfactuals.

LlamaOS: Domain Specific Chatbot for Student Queries

This project develops an open-source chatbot by fine-tuning LLaMA 2 with RAFT, RAG, and a fine-tuned retriever using LLM-generated QnA data.

Detect PIID information in written essays

Detecting personally identifiable information (PII) in student writing using Longformer & Deberta. Bronze medal in this featured Kaggle competition.

Small Language Models for ecommerce

Fine-tuned Small Language Models (SLMs) to test their efficacy against LLMs on domain-specific tasks.

Symbolic Regression using Transformers

Transformer Models for Symbolic Regression trained on a subset of Feynman Dataset. You can find it here.

Biomass Supply Chain Optimization

Biomass yield forecasting using AutoML and large-scale optimization using Mixed Integer Linear Programming combined with density-based clustering. Solution for the Shell.ai Hackathon 2023.

LLM - Detect AI Generated Text

Identifying which essay was written by a large language model.

Enabled - Community platform app for PWDs

Feature packed and completely accessible application for PwDs.

NLP Specialization

Neural networks and Deep learning

Improving deep neural networks

Structuring Machine learning projects

Convolutional Neural Networks

Fundamentals of Deep learning

Applications of AI for anomaly detection

Applications of AI for predictive maintenance