CV

Download PDF

Education

University of California, Berkeley, Master of Data Science, Machine Learning emphasis. GPA 4.0. Expected May 2027. Coursework: NLP with Deep Learning, Computer Vision, Applied Machine Learning, Causal Inference, Data Engineering, Statistics.

University of California, Riverside, B.S. Computer Science. June 2025.

Publications

1. V. Vankadaru, V. DuTell. Vision-Language Models Trust What They Can Discern. eXCV Workshop, ECCV 2026.
2. V. Vankadaru, A. Matthews, T. Roosta, P. Passban. Readable but Not Controllable: Neuron-Level Evidence for Medical LLM Hallucination. DAIH Workshop, COLM 2026. arXiv:2607.00158
3. A. Matthews, V. Vankadaru, T. Roosta, P. Passban. Rethinking Medical LLM Hallucinations: A System-Level Survey. Artificial Intelligence in Medicine, in press.
4. V. Vankadaru, V. DuTell. Believing Is Seeing: How LLaVA Models Weigh a Contradicting Image. Under review, ICLR 2027.
5. E. S. Danthinne, V. Vankadaru, et al. Potential for Age-Based Bias in AI-Informed Neurosurgical Decision-Making: A Sociodemographic Subgroup Assessment Approach for Open-Source Glioma Molecular Subtype Classifiers. In review.
6. A. Ge, V. Vankadaru, X. Liu, et al. Benchmarking Somatic Mutation Representations for Cancer Typing and HRD Prediction. In review.
7. V. Vankadaru. A Recipe for Multimodal Multi-Instance Depression Detection: Frozen Teachers, Schedule-Aware Training, and Audio Contribution Diagnostics. Under review, ACL Rolling Review.

Research Experience

UC Berkeley, Graduate Researcher

Advisor: Dr. Vasha DuTell. Apr 2026 to present.

• Studied how vision-language models (CLIP-LLaVA) weigh an image against contradicting text, applying activation patching and logit lens to measure the decision; models trust whichever input they can discern. First author, ECCV 2026.
• Extended the study to three LLaVA models and four kinds of image-text conflict, adding an evaluation on real photographs (NaturalBench) and multiple-testing correction. Full paper under review at ICLR 2027.

Advisor: Prof. Tanya Roosta. Jan 2026 to present.

• Led a neuron-level hallucination study of medical LLMs (BioMistral, Meditron, Llama 3.1): linear probes on hidden activations detect hallucination at AUROC 0.77 to 0.86, yet steering those same neurons cannot reliably suppress it. First author, COLM 2026.

UCSF M2AI Lab, Graduate Researcher. Advisor: Dr. Madhumita Sushil. Jun 2026 to present.

• Externally validated three open-source MRI classifiers on a 3,380-patient cohort and showed that errors AUROC hides are driven by age: specificity 54% for patients under 40 vs 80% for 55 and older.
• Scaled ground-truth labels from 442 to 3,380 patients with an LLM extraction pipeline over 9,377 pathology reports (F1 0.96 to 0.97), which powered subgroup tests with bootstrap CIs, permutation tests, and FDR correction.

UMSOM Tensor Lab, Graduate Researcher. Advisor: Dr. Bradley A. Maron. Jun 2026 to present.

• Designed Bio MAF v4, a 948-feature interpretable representation of a tumor’s somatic mutations, and benchmarked it against mutational signatures and a from-scratch MuAt transformer on 8,800 TCGA tumors across 20 cancer types.
• Showed that feature tables with XGBoost beat the transformer (0.738 vs 0.528 balanced accuracy) and transfer to 913 held-out tumors.

Industry Experience

DASION, Machine Learning Engineer, Jun 2025 to present. ML Engineering Intern, Sep 2021 to Jun 2025. Claremont, CA.

• Built a multimodal depression detection model on clinical interview audio and transcripts in PyTorch (wav2vec2 and RoBERTa encoders, gated-attention multi-instance pooling), reaching F1 0.87 on DAIC-WOZ with occlusion diagnostics for per-modality attribution.
• Fine-tuned BERT for clinical documentation classification and deployed models as AWS inference endpoints serving 5+ institutions.
• Trained ResNet50 transfer-learning classifiers for lung and cardiac imaging (93% diagnostic accuracy) and a bidirectional LSTM with attention for voice-based depression screening (85% accuracy).

Service

• Reviewer: InterpScience Workshop (NeurIPS 2026), Actionable Interpretability Workshop (COLM 2026), eXCV Workshop (ECCV 2026)
• Presenter, DAIH Workshop at COLM 2026; lead fellow presenter, The Tensor Lab
• BlueDot Impact Technical AI Safety course, 2026

Skills

Python, PyTorch, TransformerLens, Hugging Face, scikit-learn, XGBoost, SQL, C++, R, AWS, Docker, Linux, LaTeX