Bhushan Shah
Training LLMs | Reinforcement Learning | Deep Learning | Machine Learning
Currently Exploring
01
LLM Agents
Memory & Policy Following
- ›How do agents handle and update their memory?
- ›How well do agents follow policies set by users?
02
Interpretability
Exploring the J-Space
- ›What is happening inside large language models?
- ›What does the J-space of an LLM reveal?
03
Meta-Gaming
When LLMs Game the Game
- ›When is meta-gaming observed?
- ›How do we know an LLM is meta-gaming?
- ›What challenges does it bring?
- ›How do we deal with it?
Work Experience
AI/ML Intern
May 2026 – Aug 2026
- •Built three multi-agent LLM pipelines (debate, moderator, majority-voting) to classify harm intent of abusive accounts across 17-category taxonomy, scoring 180K+ accounts and extending detection beyond binary classification to adversary objective
- •Explored taxonomy-discovery pipelines and showed LLM-generated taxonomies were unstable; motivated fixed taxonomy approach combined with account-level behavior signals
- •Produced the first quantified breakdown of why accounts are abusive across 1M+ unrestricted accounts, enabling enforcement prioritization by scam type
- •Delivered labeled output that reduced manual review effort, supplied training data for distilling smaller open-source models, and became foundation for scammer-persona agent
LLMsAI SafetyLLM Agents
Project Intern
IBM
Sep 2024 – March 2025
- •Curated 2M+ food products labeled dataset from OpenFoodFacts with nutrition, ingredient, and health scores
- •Benchmarked GPT-4o, Mistral Large 2, Llama 3.1 405B, and Gemini Flash 2; identified GPT-4o and Mistral Large 2 as top performers for score prediction task
- •Built Generative AI-powered application allowing users to search food product scores via product name, URL, or image input
- •Optimized inference pipeline through caching, improving latency 10× (2s → 0.2s) for production deployment
- •Developed and integrated LLM inference and function-calling pipelines with backend microservices using FastAPI
WatsonXLLMsAI Applications
Tech Stack
Python
PyTorch
TensorFlow
Verl
TRL
Lighteval
Liger Kernel
LangChain
LangGraph
Scikit-learn
NLTK
Pandas
NumPy
Seaborn
Matplotlib
OpenCV
OpenAI
Hugging Face
Gym
React
TypeScript
Next.js
Docker
AWS
Highlighted Projects
Research Papers Implemented
Bias DetectionMedia BiasPLMs
Decoding News Bias: Multi-Bias Detection in News Articles
Bhushan Santosh Shah, Deven Santosh Shah, Vahida Attar
Comprehensive framework for detecting multiple types of biases in news articles using LLM-based dataset annotation and transformer-based classification.
Paper
LLM AgentsSpeculative InferenceLatency Optimization
Multi-Step Speculative Actions: From Single Guesses to Speculative Tree in LLM Agents
Bhushan Santosh Shah et al.
Accelerating LLM agents through joint breadth-depth speculation, achieving 40% latency reduction and overcoming sequential bottlenecks in agentic systems.
Paper