Research
My research spans perception, privacy, and deployment — building ML systems that work reliably in the real world. The timeline below traces how my focus has evolved across roles and institutions.
2023 – Present
Machine Learning Research Engineer · Mindtrace, Manchester
Agent-Based Systems & Distributed AI
My current focus is building autonomous agents for industrial asset inspection — distributed monitoring agents that assist weld inspection workflows, coordinate across tasks, and surface anomalies to human operators in real time.
Core research questions:
Distributed Multi-Agent Lifelong Learning — How do agents keep learning from deployment feedback across tasks without forgetting prior knowledge, and without a central server? Drawing on PEEPLL (ICONIP 2024) and DeLAMA.
Agent Memory — Short-term working context (within a session) and long-term knowledge (across sessions and deployments). Exploring structured memory architectures including A-MEM and collective memory patterns for multi-agent teams.
Agentic RAG & CAG — When should an agent retrieve vs. use cached knowledge? Investigating Agentic RAG patterns and Cache-Augmented Generation as a lower-latency alternative.
Long-Horizon Planning & Self-Reflection — Agents that decompose inspection tasks into sub-goals, re-plan when execution diverges, and iteratively critique their own outputs before acting.
Agent Safety in Production — Evaluating and constraining agents operating real tools in live industrial environments. Drawing on OpenAgentSafety and AgentMisalignment, 2025.
Key methods: Multi-agent coordination, FedAvg/FedProx, RAG, CAG, KV-cache memory, hierarchical planning
Computer Vision & 3D Deep Learning
Visual perception at industrial scale — models that understand both 2D imagery and 3D spatial data for inspection and automation.
In 2D deep learning: self-supervised learning with DINO architecture features on large-scale unlabelled datasets, parameter-efficient fine-tuning via LoRA achieving 95% accuracy on industrial inspection tasks, and detection/segmentation with YOLOv10 and DeepLab.
In 3D deep learning: point cloud segmentation with PointNet, PointNet++, and SPVCNN on LiDAR-captured infrastructure data. Built preprocessing pipelines using PDAL, GeoPandas, and Shapely for coordinate transformations, ground filtering, tiling, and shapefile integration at scale.
Key methods: DINO architectures, LoRA, YOLOv10, DeepLab, PointNet, PointNet++, SPVCNN/SPVNAS, PDAL, GeoPandas
2022 – 2023
MSc Artificial Intelligence · University of Essex
Federated Learning & Privacy
Investigated the tension between model performance and data privacy in conversational AI. Developed privacy-preserving architectures for self-learning chatbots under federated training — keeping sensitive user interactions on-device throughout the learning process.
Also applied federated learning to Human Activity Recognition on IoT devices: integrating an Empatica E4 wearable with a custom Android app, training CNN-LSTM models on-device with TensorFlow Lite, and federating updates using FedAvg, FedProx, and personalised aggregation strategies over MQTT/RabbitMQ. Final model matched centralised baselines while keeping all data on-device.
Key methods: FedAvg, FedProx, Personalised Federated Learning, TensorFlow Lite, MQTT/RabbitMQ, CNN-LSTM
Publications:
- Privacy-Preserving Self-Learning Chatbot with Federated Intelligence — IEMTRONICS 2025
- Evaluating Dialogue Adaptability: A Comparative Study of Self-Feeding Mechanisms in Federated and Centralized Chatbot Architectures — SAIA 2025
Submitted:
- MP-CFL: Matrix-Profile Clustered Federated Learning for Communication-Efficient Training under Data Heterogeneity — IEEE TNNLS (under review)
NLP & Conversational AI
MSc dissertation: attention-based neural networks for conversational agents. Used a Reformer (efficient Transformer) for multi-turn dialogue modelling with an NLU component for intent recognition, applied to a closed-domain hotel customer service scenario. This work directly motivated the federated chatbot research.
Also built sequence-to-sequence models (encoder-decoder RNN/LSTM with attention) for text-to-SQL generation, trained and evaluated on the Spider benchmark — covering nested queries, joins, and multi-column selection.
Key methods: Reformer, BERT, Seq2Seq RNN/LSTM, Attention, NLU, Spider Dataset
Medical Image Segmentation
Participated in the HuBMAP + HPA Kaggle competition (Sep 2022) — segmenting functional tissue units (FTUs) across five human organs in high-resolution microscopy images.
Implemented UneXt50 with an EfficientNet encoder, augmented with FPN for multi-scale feature fusion and ASPP between encoder and decoder. Trained with Lovász + Dice loss on TPU with checkpointing to GCS.
Result: Top 14% global ranking
Key methods: UneXt50, EfficientNet, FPN, ASPP, Lovász Loss, Dice Loss