Aravind Narayanan

Toronto, Canada

Hello! I’m Aravind Narayanan, an Associate Applied Machine Learning Specialist at the Vector Institute in Toronto (full-time since June 2025). I completed my Master of Engineering in Electrical and Computer Engineering at the University of Toronto, specializing in Data Analytics and Machine Learning, and hold a Bachelor’s in Electronics and Communication Engineering from IIIT Hyderabad.

My work at Vector sits at the intersection of explainability, multimodal AI, and agentic systems. I serve as Technical Lead for the Interpretability in LLMs and Agents Bootcamp and the Machine Learning Applications (MLA) Program, both delivered to industry sponsor companies. On the research side, my recent papers include:

AgentFinVQA — a deployable multi-agent pipeline for auditable financial chart QA (under review)
From Features to Actions — comparing attribution-based and trace-based explainability across static and agentic AI (accepted, FTC 2026)
VLDBench — a 62K-sample benchmark for multimodal disinformation detection aligned with AI governance frameworks (accepted, Information Fusion, Elsevier)

Previously, I worked at the Neural Robotics Lab on monocular depth estimation for human-robot environments, contributed to ML-driven news clustering at the Laboratory for Applied Informatics Research, and built pyMLV at the Bernhardt-Walther Lab for mid-level visual representation research.

I’m proficient in Python, C++, and SQL, with deep expertise in PyTorch, TensorFlow, and cloud platforms (GCP, AWS). Feel free to explore my work on GitHub!

news

Jun 18, 2026

Excited to share our new preprint “AgentFinVQA: A Deployable Multi-Agent Pipeline for Auditable Financial Chart QA” (with Shaina Raza), currently under review. AgentFinVQA achieves +7.68 pp over zero-shot baselines on FinMME while maintaining full auditability and on-premise deployability. Check it out on arXiv!

Jun 15, 2026

Serving as Technical Lead for the Machine Learning Applications (MLA) Program hosted by the Vector Institute for sponsor companies, delivering applied ML curriculum and technical mentorship.

Jun 01, 2026

Serving as Technical Lead for the Interpretability in LLMs and Agents Bootcamp organized by the Vector Institute for sponsor companies, covering mechanistic interpretability, explainability tools, and agentic AI diagnostics.

May 31, 2026

Our paper “From Features to Actions: Explainability in Traditional and Agentic AI Systems” has been accepted to the Future Technologies Conference (FTC) 2026! This work compares attribution-based explanations with trace-based diagnostics across static and agentic AI settings. arXiv

Project Page

Feb 28, 2025

Our paper “VLDBench: Evaluating Multimodal Disinformation with Regulatory Alignment” has been accepted at Information Fusion (Elsevier)! VLDBench is a 62K-sample benchmark for multimodal disinformation detection across 13 categories, curated from 58 news outlets and aligned with AI governance frameworks. Paper

Project

Jan 14, 2025

Currently working as an Applied AI Intern at the Vector Institute since Jan 2025, where I’m developing evaluation frameworks for multimodal large language models.

Jul 22, 2024

Presenting a poster at the 2024 Toronto Robotics Conference!

May 01, 2024

Started as a Research Assistant at the Laboratory for Applied Informatics Research (LAIR) working with Prof. Javed Mostafa

Apr 15, 2024

Started as a Computer Vision Intern at the Neural Robotics Lab working with Prof. Brokoslaw Laschowski

latest posts

Jun 23, 2026	PyTorch Lightning 101 — Are You Still Writing the Same PyTorch Boilerplate Every Single Project?
Jun 23, 2026	JAX: The Python Library That Makes PyTorch Engineers Feel Like They're Learning to Code Again

selected publications

AgentFinVQA: A Deployable Multi-Agent Pipeline for Auditable Financial Chart QA

Aravind Narayanan, and Shaina Raza

arXiv preprint arXiv:2606.19782, 2026

Under Review

Abs arXiv PDF

Financial chart question answering in regulated settings demands more than accuracy: practitioners must know which answers to trust before acting on them, and many institutions cannot send client data to external model providers. We present AgentFinVQA, a multi-agent pipeline that decomposes each query into planning, OCR, legend grounding, visual inspection, and verification, recording every step in a traceable Model Evaluation Packet (MEP) per sample. AgentFinVQA improves +7.68 pp over a zero-shot baseline on FinMME while maintaining full auditability and on-premise deployability.
From Features to Actions: Explainability in Traditional and Agentic AI Systems

Sindhuja Chaduvula, Jessee Ho, Kina Kim, and 6 more authors

2026

Abs arXiv PDF Website

Over the last decade, Explainable AI has primarily focused on interpreting individual model predictions, producing post-hoc explanations that relate inputs to outputs under a fixed decision structure. Recent advances in large language models (LLMs) have enabled agentic AI systems whose behaviour unfolds over multi-step trajectories. In this work, we bridge this gap by comparing attribution-based explanations with trace-based diagnostics across both settings. Our results show that while attribution methods achieve stable feature rankings in static settings (Spearman ρ = 0.86), they cannot be applied reliably to diagnose execution-level failures in agentic trajectories.
VLDBench: Evaluating Multimodal Disinformation with Regulatory Alignment

Shaina Raza, Ashmal Vayani, Aditya Jain, and 8 more authors

Information Fusion (Elsevier), 2025

Abs arXiv HTML PDF Website

We introduce the Vision-Language Disinformation Detection Benchmark (VLDBench), the first large-scale resource supporting both unimodal (text-only) and multimodal (text + image) disinformation detection. VLDBench comprises approximately 62,000 labeled text-image pairs across 13 categories, curated from 58 news outlets. Evaluations of state-of-the-art LLMs and VLMs show that incorporating visual cues improves detection accuracy by 5 to 35 percentage points over text-only models.