All Projects

Designed and developed a human-sized robotic mannequin for chemical exposure testing and environmental sensing at the University of Hertfordshire.

System Overview

A fully articulated robotic platform that simulates human movements and postures while collecting comprehensive environmental and exposure data through distributed chemical sensors.

Key Components

• Mechanical Design: Human-proportioned structure with multiple degrees of freedom
• Motor Control: Precision control using Maxon motor drivers for realistic motion
• Sensor Network: Distributed chemical sensors across the mannequin body
• Communication: Modbus RTU protocol across Arduino and Raspberry Pi modules
• Web Interface: Real-time monitoring and control dashboard for remote operation

Technologies

Python, Arduino, Raspberry Pi, Modbus RTU, Maxon Motors, Control Systems, Distributed Sensors, Web Interface

Applications

• Personal protective equipment (PPE) testing
• Workplace exposure assessment
• Chemical safety research
• Environmental monitoring studies
• Ergonomics and safety protocol development

Technical Achievements

• Seamless integration of heterogeneous embedded systems
• Reliable distributed sensor data collection
• Intuitive web-based control interface
• Repeatable and programmable motion sequences
• Real-time data visualization and analysis

Developed a comprehensive web-based platform for synchronised multimodal emotion data collection, integrating EEG (MUSE 2), facial emotion recognition, and human feedback for large-scale emotion studies.

Key Features

• Real-Time EEG Streaming: Integration with MUSE 2 headset for live brainwave data collection
• Facial Emotion Recognition: Vision Transformer (ViT) based emotion inference from video
• Synchronized Data Collection: Temporal alignment of EEG, facial expressions, and self-reported emotions
• Large-Scale Study Support: Designed for 250+ participants studying emotional responses to video content
• Multi-Dataset Training: ViT classifiers fine-tuned on FER2013, RAF-DB, AffectNet, and CK+ datasets

Technologies

Python, JavaScript, React, Node.js, PyTorch, MUSE 2 SDK, Vision Transformer (ViT), FER2013, RAF-DB, AffectNet, CK+

Technical Highlights

• Web-based architecture for remote participation
• Real-time data streaming and processing
• Robust emotion classification using state-of-the-art ViT models
• Comprehensive data annotation and quality control pipeline
• Secure storage and privacy-preserving data handling

Applications

The platform enables research in affective computing, emotion AI, and mental health monitoring by providing high-quality multimodal emotion datasets.

Facilitated a collaborative research sprint at the Alan Turing Institute with multidisciplinary teams, tackling industrial challenges in advanced manufacturing using data-driven approaches for AMRC (Advanced Manufacturing Research Centre).

Project Objectives

Addressing challenges in manufacturing data analysis with focus on:

• Low-frequency data problems
• Sparse dataset analysis
• Data augmentation strategies
• Predictive modeling for quality control

Contributions

• Data Augmentation: Applied advanced techniques to improve analysis robustness in sparse datasets
• Synthetic Data Generation: Developed methods for creating realistic synthetic samples
• Dimensionality Reduction: Implemented PCA and SVD for feature extraction
• Team Facilitation: Coordinated multidisciplinary collaboration and knowledge sharing
• Report Co-Authoring: Contributed to published outcomes and recommendations

Technologies

Python, Data Analysis, PCA, SVD, Data Augmentation, Synthetic Data Generation

Impact

The project provided actionable insights for AMRC partners and demonstrated effective approaches for dealing with challenging industrial datasets common in manufacturing environments.

Project Link: Alan Turing Institute - Data Study Groups

Key Takeaways

• Innovative solutions for sparse and low-frequency data challenges
• Cross-domain collaboration between academia and industry
• Practical application of advanced data science techniques
• Rapid prototyping and validation methodologies

Linguagent is a personalized, adaptive learning tutor built as a team of specialist AI agents on Google ADK and Gemini. Its flagship is language, assessing learners against the CEFR standard, planning their day, and tutoring their writing, with every judgement grounded in a real rubric. The same adaptive engine teaches any subject or skill, and learners can ground any journey — language included — on their own resources: files, links, web research, text, or AI-generated content, making each journey as specific or as generic as they like. Behind every journey is a living learner model that updates after each practice turn and shows progress over time, in the learner’s own language. Every agent is decoupled and exposed over A2A, so it’s a B2C tutor today and a service institutions can call tomorrow, all running on Cloud Run and Vertex AI.

Problem

Learning anything well at scale is either generic (one-size-fits-all courses that ignore your level, goals, and the material you actually care about) or expensive (1:1 tutoring). AI could close this gap, but most AI tutors generate ungrounded feedback — confident, but not anchored to any standard or to your own resources — and assume you already speak English to use them at all.

Solution

Linguagent is a team of specialist AI agents behind one API that run the same adaptive loop for any learner:

Assess → Score (grounded) → Plan → Tutor → Update the learner model → Show progress → Re-plan

• Assess with a tailored, interest-themed diagnostic (a quick quiz, or self-rate).
• Score answers grounded in a real rubric, producing a structured learner profile (per-skill level + confidence + cited evidence).
• Plan a daily plan weighted to the learner’s weakest, most goal-relevant skills, inside their time budget.
• Tutor the learner’s work with prioritised, rubric-anchored feedback, feeding a measurable update back into the learner model after every attempt.

Two Modes, One Engine

• Language, graded against the CEFR standard (reading, writing, grammar, vocabulary), grounded on CEFR descriptors and, optionally, the learner’s own resources.
• Subject / skill mastery, for learning any subject or skill (a certification, a course, a topic), graded on a mastery scale (novice → expert) per area.

Every journey is shaped by the learner’s own settings — what and why (target language or subject, goal, and interests that theme every question and activity), how it’s graded (quick quiz or self-rate; CEFR or mastery), where the content comes from (the journey’s Knowledge Bank: own files, web links, pasted text, or an autonomous web-research briefing — with a choice of how much to lean on it, from own-material-only to fully AI-generated), and feedback in the learner’s own language with a plan sized to the minutes they have.

Technologies

Google ADK, Gemini 2.5 Flash, Vertex AI, Cloud Run, Firestore, Firebase Authentication, Cloud Build, Artifact Registry, Google Search grounding, A2A (ADK to_a2a), MCP (Model Context Protocol) with the Tavily MCP server, FastAPI, Pydantic, Python 3.11, React + Vite + TypeScript, Zustand, Tailwind CSS, Firebase JS SDK, openapi-typescript.

Data Sources

A curated CEFR rubric corpus (A1–C2; original descriptors, swappable for licensed sets or Vertex AI Search) that agents retrieve to ground scoring. Learners’ own files, links, and text ground a journey via RAG. Live web comes from Tavily (MCP) and Google Search; Gemini 2.5 Flash (Vertex AI) handles generation. Cloud Firestore holds the learner model (skill state, history, snapshots, Knowledge Banks, usage), with Firebase Auth scoping each user’s data.

Findings

Grounding judgments in a real rubric, with cited evidence, is what makes the adaptive learner model trustworthy. In ADK, a structured-output agent is terminal (no tools), so each grounded step is a tool-using grader feeding a structured formatter. Real learning has a human mid-flow, so agents are orchestrated from the API layer rather than a single graph. Generic-vs-specific and language-vs-any-subject are just settings on one adaptive loop — scope grows, the architecture doesn’t.

Live Product

Try it: linguagent.com | Demo video: YouTube

Developing an AI-driven multimodal pipeline for ambient assistive technology in virtual wards supporting post-surgery and heart failure recovery at the University of Hertfordshire.

Key Features

• Multi-View Human Activity Recognition: Designing advanced HAR systems using multiple camera viewpoints for robust patient monitoring
• Sensor Fusion Framework: Integrating multiple sensor modalities for comprehensive health and activity assessment
• Human-Robot Interaction: Creating HRI scenarios to enhance patient engagement and provide personalized assistance
• Adaptive Assistance: Real-time monitoring and intervention system adapting to patient needs
• Healthcare Integration: Bridging healthcare and social care systems

Collaboration

• Princess Alexandra Hospital
• European research partners
• Funded by Dinwoodie Charitable Company Research Grant (2025-2027)

Technologies

Python, PyTorch, ROS2, Deep Learning, Sensor Fusion, Multi-View HAR, HRI, Ambient Intelligence, Healthcare AI

Impact

The system enables patients to recover at home with continuous AI-powered monitoring and robotic assistance, reducing hospital stays while maintaining high-quality care standards.

More info: Robotics Research Group, UH

Developing AI-driven intent-detection systems using multi-sensor biomechanical data (EMG, IMU, kinetics, kinematics) with meta-learning and Transformer architectures at the University of Hertfordshire.

Key Contributions

• Deep Learning Framework: Designed and implemented a sensor fusion framework for real-time motion recognition and adaptive human-robot interaction
• Meta-Learning: Applied meta-learning techniques for rapid adaptation to new users and movement patterns
• Transformer Architectures: Developed attention-based models for temporal pattern recognition in biosignals
• Dataset Creation: Led data collection campaigns and created benchmark datasets for wearable sensor research
• European Collaboration: Working with partners across Europe on system integration and evaluation

Technologies

Python, PyTorch, CUDA, ROS2, HRNet, YOLOv7, Meta-Learning, Transformers, Edge AI

Impact

The project aims to create assistive wearable garments that can predict user intent and provide appropriate assistance for people with mobility impairments, enabling more natural and intuitive human-robot collaboration.

More info: SWAG Project Website | Robotics Research Group, UH

Leading ongoing research in human activity recognition via multi-view camera systems, initiated during PhD and extended through postdoctoral collaborations at the University of Hertfordshire.

Dataset Contributions

RHM-HAR-SK: Multi-view skeleton-based activity recognition dataset with:

• Multiple synchronized camera viewpoints
• 3D skeleton annotations
• Diverse activities for assisted living scenarios
• Benchmark evaluation protocols

RHM-HAR-1: RGB multi-view activity recognition dataset featuring:

• High-resolution video from multiple angles
• Natural home environment recordings
• Comprehensive activity taxonomy
• Public release for research community

Technical Innovations

• Advanced AI Architectures: Integrating Transformer-based models for temporal modeling
• Agentic AI Concepts: Exploring memory-driven and goal-oriented recognition systems
• Multi-View Fusion: Novel approaches to combining information from multiple camera perspectives
• Edge Deployment: Optimizing models for real-time performance on resource-constrained devices

Technologies

PyTorch, MMPOSE, YOLOv7, GPU Cluster, Transformer, Agentic AI, Multi-View HAR

Publications

Multiple papers published on dataset design, multi-view recognition, and intelligent system architectures, contributing to the broader HAR research community.

Impact

The RHM-HAR series has become a valuable resource for researchers working on ambient assisted living, elderly care monitoring, and smart home applications.