Khashayar Ghamati | Principal AI & Robotics Engineer

About Me

I’m Khashayar -- an AI engineer working on systems that keep learning after they’re deployed.

As Principal AI Engineer at Inframatic Engineering, I lead adaptive AI for compliance automation in critical infrastructure -- knowledge-graph retrieval, multi-agent architectures, and computer vision, built to run in production rather than in a notebook. Behind that sits a decade of software engineering and a habit of making research-grade ideas reliable enough to ship.

Alongside that, I distilled a 72B language model into a 1.5B agent released on Hugging Face, and I build transformers from scratch to understand the models I optimise.

That practice is grounded in published work. I defended my doctoral thesis at the University of Hertfordshire in June 2026 -- NOVA, a framework for novelty detection and continual learning in human–robot interaction, validated on the ARI humanoid -- and I still supervise BSc and MSc students there. My work appears in IEEE TCDS, RO-MAN, ICSR and Sensors; I review for IEEE TCDS and co-organised SS03 at ICSR 2026.

If you’d like to get in touch about collaboration, speaking, or adaptive AI in real-world systems -- say hello.

Industry Experience

Principal AI Engineer

Inframatic · 2024--Present

Working on Inframatic's production generative-AI platform for compliance automation in UK rail infrastructure, including Graph RAG retrieval over large real-world document corpora.

AI Engineer

Vect AI · 2023--2024

Principal developer and technical lead for a generative-AI platform for construction businesses: designed the RAG system, built the web application, backend and the majority of the RAG pipelines, and took the platform to production, including a client deployment.

Technical Lead

SnappRoom! · 2022--2023

Product Manager

Alef Market · 2019--2021

Managed delivery of warehouse and finance software.

Senior Software Engineer

Arta Vision AVA · 2017--2019

Full-stack and DevOps engineer: front-end, back-end and deployment.

Software Developer

Karina Mobile Solutions · 2016--2017

Built mobile applications.

Earlier Roles

2011--2015

Software Developer at Kaman Academy (2014--2015) and Atie Dade Pardaz (2013--2014); Volunteer Software Developer at the Mechatronic Research Lab, MRL (2011--2013), where the RoboCup journey began

Education

Thesis: Robot Adaptation to Unseen Environments: On-the-Fly Learning Using Novelty Detection and RL

Viva passed June 2026; award recommended subject to minor amendments

Supervisors: Dr. Abolfazl Zaraki, Prof. Farshid Amirabdollahian, and Dr. Frank Foerster

Funded PhD Studentship focusing on agentic AI and continual learning for adaptive robotics

Thesis: Employing Bayesian networks to predict customers' future purchases

Distinction in machine learning and data science applications

First Class Honours with specialisation in robotics and embedded systems

Member of MRL Humanoid Kid Size and MRL At-Home Robot teams -- contributing to motion control, vision, world modeling, localization, AI, and embedded system design for RoboCup competitions

Research Interests: Continual Learning & Robotics

Agentic AI & Cognitive Architectures Continual & Lifelong Learning Novelty Detection & Transfer Learning Reinforcement Learning & Policy Transfer Human-Robot Interaction (HRI) Vision-Language Models & Multimodal AI Neuroadaptive AI & Affective Computing Explainable & Trustworthy AI

Selected Publications & Activities

Under Review

IEEE TMM

HAR-Agent: Multilingual Multimodal Activity Recognition via Knowledge-Distilled LLM Reasoning

K. Ghamati, M. R. Shahabian Alashti, A. Fallahirahmatabadi, A. Zaraki

Preprint · Model

Elsevier RAS

Inferred-Structure and In-Context Learning: A Unified Bayesian Framework for Adaptive Robotic Intelligence

K. Ghamati, A. Zaraki

DOI · Video

Elsevier ESWA

Short-Term Memory Reduces Semantic Redundancy in Vision-Language Activity Recognition

K. Ghamati, F. Amirabdollahian, F. Foerster, A. Zaraki

Preprint

2026 LLM-Augmented GraphRAG and Bayesian Networks for Cold-Start-Resilient Market Basket Recommendation

2026

ICSR 2026

Can Post-Training Adaptation Replace Continual Learning? Evidence from EEG Emotion Recognition

K. Ghamati, A. Zaraki

ICSR 2026

Vision–Language Model for Fall Detection in Socially Assistive Robotics: Zero-Shot Prompting and Few-Shot Calibration

M. R. Shahabian Alashti, K. Ghamati, A. Zaraki, B. Zhang, P. Holthaus, S. Alingal-Meethal, V. Velmurugan, G. Lakatos, A. Dickinson, F. Amirabdollahian

IEEE TCDS 2026

Learning to Gaze: Bio-Inspired Attention Adaptation Strategy for Social Robots

K. Ghamati, M. Banitalebi Dehkordi, H. Rahimi Nohooji, H. Voos, F. Amirabdollahian, A. Zaraki

DOI · Video

2025

MDPI Sensors 2025

Which AI Sees Like Us? Investigating the Cognitive Plausibility of Language and Vision Models via Eye-Tracking in Human-Robot Interaction

K. Ghamati, M. Banitalebi Dehkordi, A. Zaraki

DOI

ICSR 2025

Towards Memory-Driven Agentic AI for Human Activity Recognition

M. R. Shahabian Alashti, K. Ghamati, H. Samani, A. Zaraki

UHRA · Video

IEEE RO-MAN 2025

Cognitive Agentic AI: Probabilistic Novelty Detection for Continual Adaptation in HRI

K. Ghamati, F. Amirabdollahian, D. Resende Faria, A. Zaraki

DOI · Video

MDPI Sensors 2025

Towards AI-Powered Applications: The Development of a Personalised LLM for HRI and HCI

K. Ghamati, M. Banitalebi Dehkordi, A. Zaraki

DOI · Code

2024

IEEE Humanoids 2024

ARI Humanoid Robot Imitates Human Gaze Behaviour Using Reinforcement Learning in Real-World Environments

K. Ghamati, A. Zaraki, F. Amirabdollahian

IEEE Xplore · Video

2023

Alan Turing Institute

Identifying Tiger Stripes with Machine Learning for Environmental Investigation Agency

Data Study Group team, incl. K. Ghamati

Technical Skills

Programming

Python (10+ years) Java C++ JavaScript/React

ML frameworks

PyTorch Hugging Face Transformers Ray RLlib OpenAI Gym Isaac Sim ROS/ROS2 scikit-learn

LLM & agents

LangGraph LangChain RAG / Graph RAG Instruction tuning Knowledge distillation LoRA/PEFT Model compression Prompt optimisation Evaluation harnesses

Multimodal & vision

LLaVA-Next Qwen2.5 RGB-D EEG Eye tracking

Robotics & hardware

Kinect Arduino Raspberry Pi Embedded systems

RL & continual learning

PPO DQN/Rainbow Meta-RL Domain randomisation Sim-to-real Bandits / best-arm identification Novelty detection Bayesian inference Continual learning Transfer learning

Infrastructure & data

AWS (Bedrock, Secrets Manager, SSM, IAM) Azure AI Foundry Groq OpenRouter Docker/Compose/Portainer GitLab CI/CD Django/DRF PostgreSQL Elasticsearch Oxigraph (RDF/SPARQL) Neo4j MongoDB SQLite Vector DBs ZMQ Git Linux LaTeX Jupyter

Research & Engineering Projects

LLM-centred activity-recognition agent over a vision/audio/text harness. 17 configurations compared from 1.5B to 72B parameters (instruction tuning vs. supervised fine-tuning); 72B teacher distilled to a 1.5B student at 48× compression, running in under 1 GB of GPU memory; 89.2% on a five-language audio benchmark. Model on Hugging Face ↗

Short-term-memory mechanism aggregating LLaVA-Next frame descriptions before Qwen2.5 reasoning; cuts semantic redundancy by 72.5% across 29 activities at ~2 FPS entirely on local hardware. Project page ↗

GPT-2-class decoder-only LLM (BPE tokenisation, causal multi-head attention, KV-cache inference, pre-training, instruction fine-tuning with loss masking); MusicGen-style text-to-music transformer (EnCodec RVQ tokens, delay pattern, cross-attention over frozen T5, classifier-free guidance); LLaVA-style VLM (frozen ViT → MLP projector → LLM). Each built in verified stages to understand the architectures from first principles.

Continual adaptation system with 89% novelty detection accuracy for real-world HRI; calibrated confidence (ECE 0.062) and 80% adaptation accuracy in live sessions on the ARI humanoid

EEG-driven emotion-adaptive dialogue system with real-time neural signal processing

Eye-tracking evaluation of AI vision models against human attention patterns

RL-based humanoid robot achieving human-like gaze behaviour in natural environments

Machine learning system for tiger identification from stripe patterns

Agentic AI framework for robust human activity recognition with temporal memory -- the ICSR 2025 paper that HAR-Agent builds on. Project page ↗

Beacon-based user tracking with advanced localisation algorithms

Octomap integration with ROS for autonomous exploration in unknown environments

I'm an AI engineer, not a musician -- so I used generative AI to set classical Persian poetry (Rumi, Hafez, Ferdowsi) to blues and blues-fusion. Listen ↗

Teaching & Academic Activities

Special Session Co-Organiser, ICSR 2026

18th International Conference on Social Robotics, London (July 2026)

Co-organised SS03 – Continual Learning and Generative AI in Social Robotics with Dr Abolfazl Zaraki

MSc & BSc Project Supervisor

University of Hertfordshire (2024--Present)

Supervising final-year and research projects in AI, robotics, and machine learning -- 20+ BSc and MSc students to date

Visiting Lecturer (AI & Robotics)

University of Hertfordshire (2024--Present)

Delivering lectures on advanced AI topics, continual learning, and HRI

Workshop: From Prompt to Agent -- Build Your First Intelligent AI System

University of Hertfordshire (April 2026)

Hands-on workshop taking practitioners from prompt engineering to building agentic AI systems

RL & Robotics Workshop Instructor

University of Hertfordshire (2024)

Hands-on workshop on reinforcement learning applications in robotics

Visiting Lecturer (LLMs & Python)

Brunel University London (2023)

Guest lectures on large language models and practical AI implementation

Technical Support Tutor

University of Hertfordshire (2023)

Supporting undergraduate and postgraduate students in AI and programming modules

Awards & Recognition

Funded PhD Studentship

Scholarship

University of Hertfordshire (2023--2026)

Competitive research funding for doctoral studies in AI and robotics

1st Place, RoboCup Competition

Champion

@Home and Humanoid Robots Leagues (2013--2015)

International robotics competition focusing on autonomous robot behaviour

Contact

If you're interested in research collaboration, consulting, or simply want to connect, please reach out!

Email: [email protected]

Based in: West Berkshire, United Kingdom

Available for: Research collaborations, consulting, speaking engagements, and academic partnerships