Hi, I'm Keshav Bulia

Developing a deep learning model to predict 3-month functional outcomes (mRS) in acute ischemic stroke survivors using MRI/CT scans.

• Built preprocessing pipeline for 200+ patient scans with inconsistent formats
• Improved U-Net segmentation accuracy from 80.8% to 86.5%
• Doubled recall for poor outcomes (38% → 77%), identifying twice as many at-risk patients

Systematic evaluation of open-source web automation agents across frontend frameworks and LLM configurations.

• Benchmarked 4 agents across 5 frameworks with 2 LLM backends (240 test runs)
• Developed deterministic root cause analysis achieving 70% accuracy
• Found BrowserUse performed best (68.3%), DOM parsing was primary failure mode

Applying ML to predict quantum mechanical properties in 2D materials, enabling faster material screening for spintronic applications.

• Curated dataset of 200 2D materials via web scraping and NLP extraction
• Engineered 58 physics-based features achieving R² = 0.71
• Exploring ensemble and neural network approaches for improved accuracy

Investigating how cognitive injection strategies guide LLMs to emulate individual human writing styles.

• Collected participant data: articles, personality tests, concept mappings
• Combined cognitive features improved authorship accuracy by 20%
• Evaluated using cosine similarity and statistical significance testing