Master's student in AI and Machine Learning at Drexel University (GPA: 4.0). Currently working as a Research Assistant developing novel neural network initialization algorithms and extending BioMedBERT pretraining.
Using mechanistic interpretability to extract harmful direction vectors from Llama 3.1 8B, then training auxiliary MLPs to steer activations away from unsafe directions.
Extended Bayesian Initialization to convolutional architectures using unfold operations and product quantization. 81% on MNIST with 20% data, 65% on CIFAR-10 without gradient descent.
Extended BioMedBERT pretraining on PubMed abstracts (2020-2024) with robust BLURB evaluation (13 tasks, 10 seeds), exceeding baselines on BioASQ and PubMedQA.
I'm an AI/ML Researcher and Engineer pursuing a Master's in AI and Machine Learning at Drexel University. My research spans Bayesian neural network initialization, mechanistic interpretability for LLM safety, and biomedical NLP. Previous experience includes building enterprise RAG pipelines and deepfake systems at Ernst & Young, and backend infrastructure at Neuflo Solutions.
Drexel University, Philadelphia, PA
Expected June 2026 | Drexel CCI Merit Scholarship Recipient
Mahatma Gandhi University, India
Graduated 2022 | Best Outgoing Student
University of Kerala, India
Graduated 2020
Drexel University · Social NLP Lab
Drexel University · CODED Labs
Neuflo Solutions
Ernst & Young
Zoho Technologies
Extended Bayesian Initialization to convolutional architectures using unfold operations and product quantization. Achieved 81% accuracy on MNIST with 20% data, 65% on CIFAR-10 without gradient descent.
Used mechanistic interpretability to extract harmful direction vectors from Llama 3.1 8B's residual stream, then trained lightweight auxiliary MLPs to steer activations away from unsafe directions.
Extended BioMedBERT pretraining on PubMed abstracts (2020-2024) with robust BLURB evaluation (13 tasks, 10 seeds), exceeding published baselines on BioASQ and PubMedQA.
Evaluated 6 attack methods on ResNet-18 and YOLOv5: white-box PGD reduced accuracy from 83% to 0.58%; black-box patch attacks dropped classification by 17.85% and detection by 43%.
Multi-step agentic loop where GPT-4 uses chain-of-thought reasoning to autonomously select and invoke external tools (Wikipedia, geosearch, live Google News RSS) via JSON schema-defined interfaces.
Complete LSTM network from scratch using only NumPy — gated forward/backward passes with BPTT, custom BPE tokenizer, trained on song lyrics with top-k sampling.
Multi-class Logistic Regression with softmax and SVMs with one-vs-all classification from scratch in C++ using Eigen, benchmarked on CIFAR-10 (50K train, 10K test).
COVID-19 news video sentiment analysis using LSTM, Bi-LSTM, CNN, GRU. Published at ICITA-2021.
Ongoing research contributing to a longitudinal study of internalized self-stigma among people who use drugs on Reddit, using LLM benchmarking and prompt engineering for stigma indicator classification.
Published research on analyzing public sentiment from YouTube comments on COVID-19 news videos using deep learning models.
A deep dive into the transformer architecture and self-attention mechanism that powers modern NLP models.
A beginner-friendly introduction to PyTorch, covering tensors, autograd, and building your first neural network.
Interested in AI/ML research collaboration, consulting, or just want to chat about machine learning? I'd love to hear from you.
Built with Astro, TailwindCSS, and DaisyUI.
·
© 2026 Milan Varghese. All rights reserved.