About Me
I'm interested in problems that are just structured enough to be solvable and just unpredictable enough to stay interesting, especially in machine learning and related areas. I spend a fair amount of time working with models that don't always cooperate, trying to get them to do something useful anyway. Somewhere between debugging, experimenting, and convincing myself that the next run will finally converge, I've developed a preference for challenges that reward persistence more than elegance. It's less of a hobby and more an ongoing negotiation with complexity, and I tend to stick with it until it gives in.
Technical Skills
Languages
Frameworks & Libraries
Tools
Experience
Fairness and Data Efficiency Lab (FADE)
Undergraduate Researcher
- Currently researching the working of vision models in the field of class incremental learning (CIL).
- Implemented and reproduced results of latest state-of-the-art papers in CIL.
Machine and Medical Vision Lab (MMVL)
Undergraduate Researcher
- Applied machine learning techniques to analyze multimodal healthcare data for classification and pattern discovery.
- Worked with large-scale datasets, engineered features, and built scalable pipelines using Python and PyTorch.
- Explored dimensionality reduction and low-rank fusion methods to optimize performance and interpretability.
Research Interests
Class Incremental Learning
Investigating how vision models can continuously learn new classes without catastrophically forgetting previously acquired knowledge.
Currently WorkingReinforcement Learning from Human Feedback
Exploring how RLHF can align large language models with human intent, improving helpfulness, safety, and instruction-following.
Looking to ExploreScaling for LLMs
Interested in the systems and algorithmic challenges behind training and serving large language models efficiently.
Currently LearningProjects
A collection of research implementations, course projects, and personal explorations.
SLM-MoE: Small Language Model with Sparse Mixture of Experts
Custom 12M parameter Sparse MoE decoder-only Transformer built from scratch. Features auxiliary load-balancing loss and trained on the TinyStories dataset.
Masked Autoencoder for Self-Supervised Image Representation Learning
MAE architecture with a Vision Transformer encoder. Trained on CIFAR-10 with 75% masking, achieving 61.25% linear probing and 50.4% kNN accuracy.
Low-Rank Multimodal Fusion for Gastric Histopathology
Multimodal pipeline fusing ResNet-50 and Sentence-Transformer features. 89.8% accuracy on PatchGastricADC22 using LMF with Attention Pooling.
In-Memory Graph Based Database
Neo4j-inspired graph DB from scratch with adjacency lists, hash maps, and custom indexing. 20+ operations with O(1) node retrieval and a Python GUI.
PositionNet for Massive MIMO-OTFS Channel Estimation
PositionNet (~3800 params) for MIMO-OTFS channel estimation. Full pipeline on EVA channel model in TensorFlow/Keras with Top-K accuracy evaluation.
Publications
Selected work currently under review.
Multimodal Training for Image-Only Deployment: Feature Distillation for Report-Free Gastric Histopathology Classification
Under review at the Conference on Computer Vision and Image Processing (CVIP), 2026. Authors: Shrihari Dumbre, Bikash Santra.
Blog
Longer-form writing on things I'm learning and building.
Coming soon, first post is in the works.
Notes
Thoughts on machine learning, research, and whatever else I find interesting.
Nothing here yet, stay tuned.