Hi, I'm Ankan.

Computer Science student interested in Machine Learning, AI, and research.

ResumedownloadGitHubcodeLinkedInworkEmailmail

ABOUT

I'm a Computer Science undergraduate at Heritage Institute of Technology. I work on machine learning, computer vision, and transformer-based models. I'm particularly interested in understanding how models work, rather than only improving their performance.

I am currently interested in working on transformer circuits and representation learning, and have hands-on experience building ML models and research systems from scratch.

I am currently seeking a research internship in the field of LLMs, mechanistic interpretability, or related areas of AI research.

EXPERIENCE

Jun 2026 – Sept 2026Summer Research InternUniversity of Calcutta
  • Analyzed NIFTY50 price dynamics using statistical dependence measures.
  • Built knowledge bases for news-driven stock movement analysis by retrieving structurally similar stocks and evaluating similarity methods like sector coherence and bootstrap robustness.
Aug 2024 – PresentUndergraduate Research AssistantHeritage Institute of Technology
  • First-authored MedCompress, a modular framework for structured pruning, PTQ, and QAT across ResNet, MobileNet, EfficientNet, and DenseNet on pathology datasets, achieving up to 8× model size reduction.
  • Applied Attention U-Net with GAN-based augmentation to MRI tumour segmentation and explored SNN-based visual recognition.
Get more details herearrow_forward

PUBLICATIONS

MedCompress: A Modular Framework for Benchmarking Compression Techniques in Medical Imaginglock

Ankan Das, D. Sen, P. Bhattacharya, S. Sarkar, D. Sengupta.

DABCon 2026 — Accepted for Oral Presentation

FEATURED PROJECTS

LiteGPT
PythonPyTorch

Implemented a GPT-style language model from scratch in PyTorch, including causal self-attention, token embeddings, and autoregressive generation.

GitHub
Crucible
Next.jsFastAPIAI/ML

An AI co-founder platform that helps hackathon teams generate, refine, and stress-test project ideas against technical feasibility and constraints.

GitHubLive Demo
Multimodal Knowledge Retrieval
LangChainQdrantVLMs

Built a multimodal document retrieval system using Docling for PDFs and Qdrant, integrating vision-language models to extract semantic info from images and tables.

GitHub
BPE Tokenizer
PythonNumPy

Implemented Byte Pair Encoding from scratch, covering vocabulary construction, pair-frequency computation, and merge learning for LLM training.

GitHubLive Demo

View more on GitHubarrow_forward

RESEARCH

  • Large Language Models
  • Mechanistic Interpretability
  • Transformer Circuits
  • Representation Learning
  • Generative AI
  • Multimodal Learning
  • Model Compression

LATEST WRITING

Coming soon...

CONTACT

Have a project, research opportunity, or just want to talk?

ResumeGitHubLinkedInEmail