RESEARCH & APPLIED WORK
Where the practice comes from
TaoQ AI's engagement work draws on a body of open-source contributions, applied AI research, and academic specialisation in machine learning and AI security. This page collects that work in one place. For client engagement examples, see Case studies.
Open-source
MAINTAINER
Ziran
Open-source AI agent red-teaming framework. Tool chain analysis, multi-phase trust exploitation, multi-agent coordination testing, and remote agent scanning over REST, OpenAI, MCP, and A2A protocols.
CONTRIBUTOR
NVIDIA Garak
Contributions to the Garak LLM vulnerability scanner, the open-source tool used to probe LLMs for prompt injection, jailbreak susceptibility, and other failure modes.
CONTRIBUTOR
Hugging Face Transformers
Contributions to the Transformers library, the most widely used open-source framework for working with large language models.
Applied research
Selected research and prototype work delivered through TaoQ AI. Some details are sanitised for client confidentiality.
Synthetic-data and temporal-detection platform for online romance fraud
Applied AI researcher leading the platform architecture, the synthetic-data pipeline, and the temporal-detection framework.
Business case
Existing romance scam detection tools achieve close to zero true-positive rates because they scan for malicious content while scam messages contain emotional manipulation, and no commercial training datasets exist for the conversation patterns that matter.
Approach
A hexagonal-architecture platform combining a configurable synthetic-conversation pipeline grounded in published criminological research on scam progression, and a temporal behavioural detection service that fuses rule-based pattern checks, statistical conversation-dynamics analysis, and ML classification trained on the synthetic corpus.
Deliverables
Vendor-neutral, multi-provider LLM stack with API-first scoring output, fully synthetic training data (no victim data), and a research-collaboration framing aimed at academic partners in fraud and LLM safety.
Independent feasibility review of an AI model layer for radar-based elderly-care monitoring
Independent technical advisor engaged through a Dutch tech-industry business developer to deliver an arm's-length feasibility review of the proposed AI model layer ahead of a public research-loan decision.
Business case
An early-stage Dutch healthtech team was preparing a public research-loan application for a 60GHz mmWave radar system intended for elderly home monitoring. The funder needed independent evidence that the AI components were technically demonstrable inside a Proof-of-Concept timeline before committing capital.
Approach
Assessed only the ML/AI layer (fall detection, gait trend prediction, person identification, edge inference, privacy and data architecture), separated from hardware and software-engineering decisions. Reframed the system as activity-monitoring and safety-alerting rather than diagnostic, to keep it out of EU MDR Class IIa device territory, and re-scoped the PoC around what was demonstrable from published evidence.
Deliverables
Confidential feasibility report rating each AI component low, medium, or high risk; a focused PoC validation plan with measurable acceptance criteria; a privacy-by-design data architecture aligned to GDPR and Dutch AVG; and a phased Phase-2 and Phase-3 extension roadmap for post-funding work.
Academic
MSc Artificial Intelligence (Distinction), University of Bath
2022 to 2025
Technical Pathway: Machine Learning, Reinforcement Learning, Robotics and Machine Vision.
Dissertation
Evaluating the Feasibility of MITRE-Aware Large Language Models for Tactics, Techniques, and Procedures Detection in Cyber Threat Intelligence Reports
Asks whether a domain-adapted Large Language Model fine-tuned against the MITRE ATT&CK framework can automate Tactics, Techniques, and Procedures (TTP) detection in unstructured Cyber Threat Intelligence reports, where manual parsing currently leads to inconsistency and missed threats. Combines causal-language-model fine-tuning over a transformer baseline with a two-stage training strategy and a hybrid evaluation regime (ROUGE plus LLM-as-judge qualitative analysis). Finds that domain-specific fine-tuning improves precision on shorter text segments but trades coherence on longer outputs and produces mixed results when mapping to specific ATT&CK techniques, pointing to hybrid approaches with knowledge graphs and continual learning as the more promising path.
Research collaboration enquiries
Open to academic partners, advisory boards, and joint research on synthetic data, agent red-teaming, and AI Act conformity engineering.
Get in touch