Light Convolutional Neural Network for Detecting COPD: Multicenter Model Development and External Validation
medRxiv (under review at The Lancet Digital Health)
I turn hard ML problems into ships — medical imaging deployed across 21 hospitals, LLMs serving 30,000 users a day.
PhD dissertation: Precise Segmentation of Thin, Elongated, and Underrepresented Structures (UAB, 2026) · CS Outstanding PhD Student Award · Advisor: Dr. Chengcui Zhang
medRxiv (under review at The Lancet Digital Health)
Manuscript in preparation
Manuscript in preparation
IEEE/CVF WACV Workshops, 2024
EMNLP Industry Track, 2023
RTIP2R, 2018
Shipped systems that run in production today — government platforms, banking AI, and a RAG pipeline acquired by Promise.
Selected systems, with simplified workflows and reported evaluation results.
Fine-tuned Qwen2.5 VLM using LoRA, FP8 quantization, and Flash Attention 2, served with vLLM on a single NVIDIA L40S GPU.
Conceptual flow only; platform request-handling components are not specified here.
Fine-tuned the vision-language model with LoRA and optimized inference using FP8 quantization and Flash Attention 2. Deployed model serving through vLLM on one NVIDIA L40S GPU.
30,000 daily active users and 1–2K peak requests per second are platform-wide figures. They describe platform traffic, not GPU inference throughput.
Medical imaging · Multicenter evaluation
COPD detection from chest CT
A lightweight 3D CNN evaluated using COPDGene development data and external cohorts.
Simplified study flow; this is not a deployment diagram or an exact model architecture.
Led architecture design and multicenter evaluation for COPDxNet.
13,043 COPDGene CT scans from 21 centers. AUC: 0.92 standard-dose and 0.88 low-dose COPDGene scans; external AUC: 0.92 on SPIROMICS and 0.82 on NLST.
Manuscript under review at The Lancet Digital Health.
NIH R01: Fissure Integrity Score — quantitative pulmonary fissure metric for COPD phenotyping. Lead developer of COPDxNet and STRIPE.
Led a 13-engineer team. Qwen2.5 VLM on Kagoj.ai (vLLM, L40S). Event-driven microservices on Kafka. Dorsal.fyi RAG pipeline.
PhD: Precise Segmentation of Thin, Elongated, and Underrepresented Structures. COPDxNet, STRIPE, THREAD, dCrack61K. NSF SaTC: ballot tabulation.
Borno (ocr.bangla.gov.bd). Triton + ONNX: 18.6× speedup. Grew the team from 3 to 10+, 10K+ daily requests, 99.5% uptime.
Taught AI and AI Lab. Shifted curriculum to project-based learning.
Research collaboration, AI systems work, or just a conversation.