Hello, I'm
Ali Akbar
AI/ML Researcher
Building intelligent systems — RAG pipelines, LLM-powered agents, and multimodal AI. Researching at the intersection of Machine Learning, Computer Vision, and Healthcare AI.
Hello, I'm
Building intelligent systems — RAG pipelines, LLM-powered agents, and multimodal AI. Researching at the intersection of Machine Learning, Computer Vision, and Healthcare AI.
I am an AI/ML researcher with hands-on experience building RAG pipelines, LLM-powered agents, and multimodal AI systems using LangChain, LangGraph, LangSmith, PyTorch, and vector databases. I completed my B.Sc. in Electrical and Electronic Engineering from the University of Dhaka and worked as a Research Assistant at the Data & Design Lab (CSE, University of Dhaka), contributing to a government-funded ML deployment project, an LLM-powered domain chatbot, and a real-time data visualization dashboard. My research spans Machine Learning, Computer Vision, and Healthcare AI across 4 published and ongoing works — including microbacterial instance segmentation, gut microbiome-based IBS diagnosis, and plant leaf disease VQA — with an acknowledged contribution on an ACL Findings 2025 paper. I was Runner-up (Top 2) at the ACM UbiComp 2021 International Nurse Care Activity Recognition Challenge. I am currently seeking opportunities in AI/ML Engineering roles at the intersection of Agentic AI, Multimodal Systems, and Healthcare AI.
B.Sc. in Electrical & Electronic Engineering
University of Dhaka (Jan 2019 – Apr 2024)
CGPA: 3.52 / 4.00
To leverage expertise in AI/ML engineering and research to build impactful intelligent systems, advance the state of the art in NLP and Computer Vision, and contribute to open-source AI.
End-to-end pipeline for processing handwritten scanned documents — extracting text via OCR, chunking semantically, and indexing into a FAISS vector store for retrieval. RAG-based QA with source references and Human-in-the-Loop feedback.
Multi-turn conversational AI chatbot using LangGraph's StateGraph and MemorySaver for graph-based state management. Real-time token streaming via Groq inference backend with Streamlit frontend.
RAG pipeline that ingests content from up to 3 simultaneous URLs and multiple PDF documents. Processes web and PDF content with HuggingFace embeddings into FAISS vector store and generates source-referenced answers with LangSmith latency tracking.
Engineered time-frequency features from raw accelerometer signals using FFT and DWT. Applied Variance Threshold, ANOVA, and Chi-square feature selection. Achieved 87% accuracy using LightGBM — contributed to runner-up at ACM UbiComp 2021 Challenge.
Max-ViT backbone-based Mask R-CNN for end-to-end microbacterial instance segmentation on M-ROSE dataset. Outperformed ResNet50 Mask R-CNN, Mask2Former, and YOLOv11 — mAP 79.31%, F1 70.01%. Applied ROI Activation CAM for interpretability.
Curated hybrid dataset of 1,109 gut microbiome samples (NCBI + Bangladeshi samples). Processed through CZID pipeline, optimized with Optuna Bayesian optimization under 5-fold CV. Logistic Regression achieved 79.73% accuracy, MCC 0.5876. Manuscript in preparation.
Contributed large-scale data collection for PlantInquiryVQA — 24,964 images, 138,078 QA pairs with reasoning chains. Evaluated SOTA MLLMs (Qwen, Grok, Mistral, Gemma). Published in ACL Findings 2025 as acknowledged contributor.
icddr,b (International Centre for Diarrhoeal Disease Research, Bangladesh)
Developing a Linux-based shotgun metagenomics pipeline using SRA Toolkit, KneadData, and MetaPhlAn to acquire, preprocess, and generate taxonomic abundance profiles from publicly available healthy-control gut mi- crobiome sequencing data. Integrating these data with organization's existing cohort of 67 IBD patient samples to support comparative microbiome analysis and downstream machine learning studies for identifying disease-associated microbial signatures.
Data & Design Lab, Department of CSE, University of Dhaka
Collected real-time field data from substations, developed data analytics solutions, preprocessed and transformed electricity-sector datasets into structured formats for system integration and data-driven decision-making, and contributed to the development of Biddyut Bandhu, a domain-specific chatbot.
Adjunct Proceedings of the 2021 ACM International Joint Conference on Pervasive and Ubiquitous Computing (UbiComp 2021), pp. 434-439, 2021.
ACL Findings 2025
Third Nurse Care Activity Recognition Challenge, 2021
HASCA Workshop, ACM UbiComp