Research
Research Interests:
Large language models (post-training and evaluation), speech, audio and music AI, multimodal deep learning, code generation, computer vision, video understanding and generation, pattern recognition, and recognition of graphical structures.
Current work: Frontier LLMs at Meta (Applied AI)
Since June 2026, I have been a Research Scientist in Meta’s Applied AI (AAI) org, working on the post-training of Meta’s frontier large language models, the Muse family. I contribute across three areas:
- LLM research. Post-training, data curation, and evaluation methods, including reinforcement learning and fine-tuning, that improve the models’ reasoning, instruction-following, and multi-turn dialogue.
- Speech, audio and music. My current focus: advancing how the models understand and generate speech, audio, and music.
- Coding. Designing expert-level tasks, benchmarks, and training data that measure capability gaps and strengthen the models’ software engineering abilities.
A common thread is rigorous measurement: building evaluations that expose where capable models still fail, and using those results to improve them.
Previous work at Meta: Video AI (Media Foundation Video team)
From July 2025 to May 2026, I designed transformer- and diffusion-based models for video enhancement (super-resolution, restoration, denoising), compression, quality optimization, and accessibility across Meta’s Family of Apps (Facebook, Instagram, WhatsApp, Messenger), and collaborated with the Meta Superintelligence Lab on large vision models (LVMs) for video generation, curation, and multimodal understanding.
Speech AI at Amazon (Alexa Speaker Understanding)
As an Applied Scientist Intern (2022), I improved Alexa speaker identification accuracy by 5% with semi-supervised learning while cutting training time by 80%, and built clustering-based annotation pipelines that labeled over 10 million hours of speech data.
Ph.D. research (RIT, 2020 – 2025)
My doctoral research at the Document and Pattern Recognition Lab (DPRL), advised by Dr. Richard Zanibbi, developed fast, interpretable visual parsers for math and chemical formulas. Using graph attention-based task interaction, I enhanced contextual information while keeping a natural and interpretable graph representation, recognizing complex notation across born-digital PDFs, typeset images, and handwritten strokes. Dissertation: Parsing of Math Formulas and Chemical Diagrams using Graph-Based Representation and Attention Models.
Ph.D. projects
Multimodal Chemical Search, A system for searching chemical reactions, molecular structures, and text in scientific literature. It integrates text, SMILES, and reaction-based queries, linking extracted reaction details with molecular diagrams and textual descriptions. The interface provides structured reaction and molecule cards for easy navigation and retrieval, supporting chemists in literature exploration and data extraction. Code is open-source and available here.
- ChemScraper, a fast and accurate molecule diagram parser using characters and graphics extracted from born-digital (vector) PDF images—without the need for OCR, GPU, or vectorization. It uses these outputs to create training data for a new approach to visual parsing of molecule diagrams in raster images (i.e., pixel-based formats like PNGs) using a multi-task, segmentation-aware convolutional neural network (CNN). Code is open-source and available here.
- MathDeck project, a system for searching PDF documents in a portion of the ACL Anthology, incorporating both formulas and text, displaying matched words and formulas in context. Its user-friendly interface includes formula ‘chips’ for easy formula creation, search, reuse and annotation. MathDeck supports both LaTeX and visual formula editing.
- Built a new open-source math formula extraction pipeline for PDF files
- Adopted distributed parallelization methods with multiple GPUs and implemented custom dataloader with dynamic batch size to fully utilize the GPU, which increased the speed of the math formula parser by 6 times
- Built new tools for visualization and evaluation of parsing results and fine-grained errors analysis
- Worked on a PDF symbol extractor, called SymbolScraper that identifies precise bounding box locations in born-digital PDF documents
- Wrote an API for recognizing handwritten and typeset formulas and output the corresponding LATEX and MathML