Text, Pixels, or Both? Evaluating Input Representations for Multimodal Document QA
Comparing text, page images, and combined inputs for document QA, including their accuracy, latency, and cost trade-offs.
Publications
Selected
research.
Document understanding, retrieval, and measuring what matters.
2026
Preserving What Matters: Semantic Scaffolds Beyond Saturation in Summarization Evaluation
Evaluating summary quality through preservation of facts, questions, and entity attributes when conventional scores saturate.
Splitting Documents at Lower Cost: Multi-Split Boundary Decisions for LLM-Based Page Stream Segmentation
Predicting multiple document boundaries in one model call to improve the accuracy–efficiency trade-off in page stream segmentation.
Earlier work
Bridging the Modality Gap: Enhancing Document Retrieval with Multimodal Embeddings
Master’s thesis at MILA / Université de Montréal on multimodal retrieval for document intelligence.
Let's Evaluate Step-by-Step: A Robust Evaluation Method of LLM Applications
A framework for more reliable evaluation of applied LLM systems, submitted to the 2nd Workshop on Generative AI and Law (GenLaw ’24).
Robots in the Middle: Evaluating LLMs in Dispute Resolution
LLM behavior in mediation and AI-based dispute resolution.
Leveraging AI for Natural Disaster Management: Takeaways from the Moroccan Earthquake
Applied AI analysis for humanitarian and crisis-response contexts.
Efficient Detection of Disguised Faces from Low-Quality Surveillance Footage
Computer vision work under low-quality surveillance conditions.