Research & Projects
Foundation models and agentic AI for biomedical systems, real-time multimodal sensing and edge inference, and large-scale healthcare data curation — built end to end and evaluated for deployment. Select a project to view details.
Agentic AI & LLM Systems for Regulated Domains — Ubicomply.ai
Ubicomply.ai LLM and agentic-AI infrastructure — SAST, DAST, and policy-analysis
pipelines served on AWS g4dn GPU instances.
As Founding Research Director, I lead HPC infrastructure and agentic-AI engineering teams (~10 engineers and developers) building production systems that reason over clinical and regulatory corpora, with the evaluation and governance layer required to deploy them responsibly.
- Shipped the Ubicomply.ai LLM and agentic-AI infrastructure, deployed on AWS
g4dnGPU instances — powering SAST (static application security testing), DAST (dynamic application security testing), and policy analysis across a multi-regulatory framework, mapping findings to HIPAA, the NIST AI Risk Management Framework, and related controls in a single assessment pipeline. - Fine-tuned open-weight LLMs (Llama, Gemma, GPT-OSS) with LoRA/QLoRA/PEFT; engineered RAG pipelines and multi-agent orchestration for tool-augmented reasoning.
- Shipped malURLBox, an LLM-assisted dynamic-sandbox analyzer at 97.3% accuracy, now distributed as a Microsoft Outlook Extension.
Mobile BCI for Game-based Motor Training and Neurorehabilitation
A volunteer wearing EMOTIV Flex 2 EEG headset controlling our mBCI maze game in iPad in real-time.
Mobile BCI with a low-latency closed loop between wearable EEG and a game engine. Built on AsyncDCNN, a coherence-based neuromorphic model with cross-session and cross-subject transfer for robust decoding of motor intent. AsyncDCNN introduces a coherence-based perceptron that models stable synaptic states with graded updates from pre- and post-synaptic activity. This supports asynchronous, parallel learning across forward, backward, and lateral pathways for reliable real-time decoding under non-stationary EEG.
SMT-Learner — Self-Supervised Foundation Model for Physiological Time-Series
A transformer auto-encoder trained with masked reconstruction on ~16,320 human and ~23,639 non-human-primate movement trajectories, learning transferable representations of motor behavior under sparse supervision. From these representations I derive a novel Exploration–Exploitation biomarker that separates early- from late-phase motor skill learning — a construct long theorized in motor neuroscience but difficult to measure outside a laboratory.
- Masked-reconstruction pretraining over spatiotemporal movement trajectories.
- Evaluation via frozen-encoder linear probing, cross-cohort and cross-species transfer, and systematic ablations on masking ratio and cohort size.
- Cross-species transfer as evidence that representations capture motor-control structure rather than cohort-specific artifacts.
mHealth ecosystem (motorlearning.games) for pediatric motor assessment and learning analytics
iPad-based neuromotor assessment for upper-limb performance in children. Integrates games, EEG, and IMU sensors to quantify spatiotemporal movement trajectories and learning dynamics.
- Developed gamified mHealth tests for upper-limb assessment
- Integrated VICON, EEG, and IMU sensors for multimodal motor analytics
- Built AWS-based data storage and analysis pipeline
- A Deep Learning Framework for Assessing Pediatric Upper-limb Motor Control and Gamified Mobile Rehabilitation. IEEE Transactions on Neural Systems and Rehabilitation Engineering (Under Review).
- Gamified mHealth System for Evaluating Upper Limb Motor Performance in Children: JMIR Serious Games, 2025.
- See all publications →
HIPAAChecker.health - Healthcare Security and Privacy Assessment
Security and privacy assessment platform for mHealth apps using static + dynamic analysis and HIPAA-aligned risk scoring.
- HIPAA threat model with 272 security patterns; validated on an annotated dataset of 78 healthcare apps with expert review
- HIPAAScoring risk model integrating CVSS v4.0 and NIST SP 800-66r2
- Static/dynamic analysis pipelines and HIPAA LLM API infrastructure deployed on AWS
National Clinical Data Warehouse (NCDW) Bangladesh
Privacy-preserving national clinical data warehouse integrating de-identified records from 220 healthcare systems — ETL, harmonization to OMOP CDM/FHIR/LOINC/SNOMED CT, and ML for vocabulary standardization.
- Wrapper API-based data integration system
- ETL pipeline and Central Data Repository
- Data marts and OLAP Engine implementation
NBA-Search Analytics
Open-source NBA analytics platform. I contributed clustering and ML methods for player performance and team analytics.
- Implemented clustering algorithms (KNN, K-Means)
- Developed player similarity analysis features
- Created advanced statistical comparison tools
vInternship.org - Supporting Blended Learning Environment Connecting Interns, Universities and Industries
Virtual internship platform connecting universities, industry, and students — structured remote projects with benchmarked team performance.
- Designed platform architecture and virtual internship protocols
- Led and mentored a team of five developers
- Conducted comprehensive research and optimization studies
Robotic Process Automation (RPA) to Automate Banking Operations
Robotic Process Automation for Bank Asia Ltd — automates repetitive core-banking workflows to cut manual error and processing time.
- Designed automation system for Bank Asia Ltd operations
- Implemented client requirements using Agile Scrum methodology
- Developed clean, robust, and reusable automation modules
Digital Document Archive System for Planning Commission of Bangladesh Government
Digital archival system for the Planning Commission of Bangladesh — digitizes and manages historical documents with Bengali OCR.
- Designed and developed comprehensive ETL pipeline
- Conducted R&D for Bengali Optical Character Recognition
- Collaborated with cross-functional development teams