Research Profile
I work at the intersection of Artificial Intelligence, Smart Computing, and Human-Centric Technologies, turning computational models into systems that run in the real world. My output spans six domains: multimodal affective computing, computer vision, applied security and cryptography, IoT and embedded systems, explainable machine learning, and interactive developer tools. The common thread is building methods that are accurate, efficient enough to deploy on modest hardware, and transparent about how they reach a decision.
Papers, preprints, and thesis work spanning six domains
Affective AI, vision, security, IoT, ML, and systems
Explainable XGBoost crop recommender on a held-out split
Every paper linked to a preprint or public source
ResearchGate
View publications and research impact
Google Scholar
Track citations and academic output
Core Research Pillars
My research program spans six domains, each grounded in published work and bridging academic theory with deployable, real-world systems.
Affective Computing & Multimodal Emotion AI
Models that recognize, track, and adapt to human emotional dynamics. This covers the Emotional State Engine (ESE) over continuous Valence-Arousal-Dominance space, the streaming Cross-Modal Transformer, long-term affective memory for LLMs, and the tri-modal OlfACT framework that adds an olfactory channel.
Computer Vision & Visual Recognition
Real-time and explainable vision systems: hybrid ORB plus template-matching logo detection (LOGOSCOPE), DeepFace age, gender, and emotion analysis, Haar Cascade with ResNet-50 face detection, and transfer learning on VGG16 and EfficientNet for digit and chest X-ray classification.
Applied Security & Cryptography
From hybrid Isolation Forest and XGBoost fraud detection under extreme class imbalance to AES-GCM and RSA-OAEP encryption with steganography (Cipher Shield), Nmap-driven network auditing (NetAudit), AI-assisted ransomware forensics (Securina), and a BIP39 and ECDSA blockchain wallet.
IoT & Embedded Systems
Decoupled ESP32 and MQTT environmental monitoring (Smart Room), browser-to-hardware control over Web Bluetooth (Zarduino), and empirical field study of AI and IoT behavioral impacts on patients and clinicians in African and Asian healthcare settings.
Explainable & Efficient ML
Decisions a user can trust and audit: SHAP feature attribution for crop recommendation (AgroTech, 98.6% accuracy), Grad-CAM saliency for pneumonia screening, OCR-to-LLM quiz generation (Smart Scan and Solve), and the multimodal Yin-AI conversational agent.
Interactive Systems & Developer Tools
Tools that make hard concepts usable: an interactive linear-algebra workbench with decomposition visualization (Ze Matrix), a trie-based real-time autocomplete benchmarked to microsecond lookups, and a privacy-first full-stack social platform with measured load performance.
Publications & Papers
A selected library of peer-reviewed papers, preprints, and thesis work. Filter by domain, open any card for the full abstract, and read the paper or try a live demo where one exists.
Research Methodology & Toolkit
A rigorous, multi-layered research toolkit combining state-of-the-art computational frameworks with mathematical and statistical validation.
Deep Learning Frameworks
NLP, Vision & LLMs
Explainability & Validation
Data Science & Imbalanced Learning
Security & Cryptography
Edge, IoT & Deployment
Active Pipeline & Current Work
Ongoing research projects currently undergoing implementation, empirical validation, or peer review.
Preprint Rollout: Twenty Manuscripts to ResearchGate
Twenty completed manuscripts spanning all six research domains are being formatted to a single-column journal style and released as open-access preprints on ResearchGate. As each goes live, its "Read Paper" link in the library above is activated, so the publication record stays current from one source of truth.
Scaling the OlfACT Tri-Modal Emotion Study Beyond the Pilot
The OlfACT framework added a volatile organic compound (VOC) olfactory channel to acoustic and linguistic streams and was validated in a controlled 15-participant pilot. Current work extends that pilot toward a larger cohort and a wider set of industrial odor conditions, with the goal of quantifying how much the olfactory channel improves affect recognition when speech is ambiguous or suppressed.
Unified On-Device Empathetic Agent
Integrating two published components into one deployable system: the streaming Cross-Modal Transformer for sub-150ms emotion recognition and the long-term affective memory vector store for personalized empathetic dialogue. The aim is a single agent that reads emotion in real time and remembers a user's emotional history across sessions, while keeping inference within a CPU latency budget.