First author of an ICML 2026 paper and Machine Learning PhD researcher at the National University of Singapore, developing reinforcement-learning post-training methods for generative models and language-model reasoning. Research spans policy optimization, reward-proportional generation, multimodal generation, and diversity/mode-collapse mitigation.
Education
Aug 2024–Jul 2029 (expected)
PhD, Machine Learning
National University of Singapore, Singapore
- Thesis: Diversity and Creativity in Generative AI
- Supervisor: Dr. Dianbo Liu
Aug 2016–May 2021
Dual Degree (Bachelor's & Master's), Biotechnology
Indian Institute of Technology Madras, Chennai, India
Experience
Feb 2024–present
Research Assistant
Artificial Scientific Intelligence Lab, National University of Singapore, Singapore
- Identified and characterized frequency amplification in expected-return optimization, then developed Inverse Probability Scaling (IPS) to recover reward-proportional terminal distributions under the analyzed dynamics; resulted in a first-author ICML 2026 paper.
- Developed and implemented RL post-training in PyTorch/Hugging Face for set-valued causal, spatial, and symbolic reasoning with Gemma-3 4B; improved valid-outcome recovery from 16.0% to 43.9%, 31.6% to 90.0%, and 9.5% to 60.7% across experimental settings.
- Designed multimodal RL pipelines in which Qwen3-0.6B/GPT-2 policies conditioned a frozen Stable Diffusion v1.4 generator; developed and evaluated CLIP/DINOv2-based density-aware reward scaling, improving Image Vendi score by 32–161% relative to GRPO across settings.
- Extended the research program to larger outcome spaces through MIPS (under review at ICLR 2027), using trajectory-based marginal-probability estimation for massive discrete spaces, and DARS (under review at AAAI 2027), using density-aware reward scaling for continuous spaces.
- Co-developed a mechanistic-interpretability study using CLIP ViT-L/14, sparse Matryoshka Transcoders, and LMM-based feature interpretation; evaluated physical-plausibility failure modes over 5,000 images per model.
- Integrated density-aware reward scaling with GRPO for multimodal navigation and LLM-based text-to-image prompt adaptation.
- Developed methods for LLM post-training and alignment, focused on reward design, diverse generation and reasoning-oriented evaluation.
- Built an RL agent that edits discrete VQ-VAE latent codes using classifier-confidence and discriminator-realism feedback.
Aug 2023–Dec 2023
Machine Learning Engineer
VAO Labs, San Francisco, CA, USA
- Built LayoutLM and OCR document-understanding pipelines for structured information extraction.
- Fine-tuned T5-family models for document question answering and answer retrieval.
- Built Llama-based applications for querying large enterprise datasets in natural language.
Apr 2022–Jul 2023
Research Assistant
Computational Neuroscience Lab, IIT Madras, Chennai, India
- Designed a brain-inspired visual attention model that learns saccadic image-scanning policies for object counting with Q-learning and recurrent neural networks.
- Co-invented a System for Monitoring Parkinson's Disease; patent granted to IIT Madras.
- Developed computer-vision and reinforcement-learning systems for neurological healthcare applications.
Jun 2021–Dec 2021
Software Developer
IIFL, Mumbai, India
- Developed full-stack software with ASP.NET, SQL Server, Angular, JavaScript and jQuery.
- Integrated, tested and documented changes using TFS and Azure in an agile workflow.
Publications
Expected Return Causes Outcome-Level Mode Collapse in Reinforcement Learning and How to Fix It with Inverse Probability Scaling
International Conference on Machine Learning, 2026
How Does My Model Fail? Automatic Identification and Interpretation of Physical Plausibility Failure Modes with Matryoshka Transcoders
CVPR Workshop on Explainable AI for Computer Vision (XAI4CV), 2026 Spotlight
Brain-Inspired Attention Model for Object Counting
International Conference on Neural Information Processing, 2023
MIPS: Mitigating Mode Collapse in Reinforcement Learning for Large Outcome Spaces
Under review at ICLR 2027
Patents
A System for Monitoring Parkinson's Disease
Indian patent, granted to IIT Madras. Co-inventor. Certificate
Skills
- Programming
- Python, SQL, C/C++, JavaScript, R, Bash
- ML & generative modeling
- PyTorch, TensorFlow, Hugging Face Transformers, Stable Diffusion, CLIP, DINOv2, VQ-VAE, latent-variable models, representation learning
- RL & LLMs
- Policy-gradient RL, GRPO, reward design/scaling, RLHF, supervised fine-tuning, instruction tuning, LoRA/PEFT, 4-bit NF4, bfloat16, alignment, reasoning evaluation; Gemma, Qwen, Llama, GPT-2
- Mechanistic interpretability
- Sparse autoencoders, Matryoshka Transcoders
- Applied AI
- Stable-Baselines, RLlib, scikit-learn, OpenCV, FastAPI, Flask, OCR, document understanding
- Experimentation & MLOps
- Weights & Biases, MLflow, distributed training, Git, Docker, Linux, Google Cloud Platform