LinkedIn | GitHub | GoodReads | Email | CV
Blog
Commentary
I am currently a machine learning engineer at NY-based startup Dime. I do research in collaboration with Dr. Md Mofijul Islam, Aman Chadha, and CCDS.
Research interests: Multilingual and Low-Resource NLP, Multimodal NLP, Document AI, LLM Benchmarking
*Actively looking for PhD opportunities (Spring/Fall 2027)
Publications & Preprints
- BaFCo: A Document Understanding Benchmark for Complex Bangla Form Comprehension - Azad et al. (ECCV 2026) [Paper | Code]
- Khondo: A Multimodal Benchmark for Document Packet Splitting of Bangla Forms - Azad et al. (Under review WACV 2027) [Paper | Code]
Ongoing Research
- Khondo v2: Document Packet Splitting for Low-Resource South Asian Forms [Extension of `Khondo']
- EvoSynth: An Evolutionary Algorithm Guided Approach for Diverse and High Quality LLM-Based Synthetic Data Generation [Idea Pitch]
- Shupto-Probad: A Mechanistic Inspection of LLM Circuits for Bangla Idioms [Idea Pitch]
Education
- MS in Computer Science, Wichita State University, 2023-2024
- BS in Computer Science & Engineering, University of Dhaka, 2016-2019
Work Experience
Founding ML Engineer, Dime April 2020 - Dec 2022 | Nov 2024 - Present
- Designed and built the production pipeline for VLM-guided US tax-document analysis: multiform parsing across 1040s, Schedule C–E, and 1099 variants, rulebook-driven PII redaction, and inter-form cross-verification.
- Built the underlying document-extraction layer: efficient bulk document upload, 1099 subtype classifier, retrieval-augmented chat over documents, and VLM-guided PII-masked bank statement transaction extraction.
- Built profile-matched knowledge retrieval and agentic orchestration: a planner delegating to DB/file sub-agents, read-only introspection agents, and MCP servers.
- Developed core fintech backend services and conversational finance-NLP systems (Dialogflow intents, predictive spend/budget features).