Agentic systems | Coding-agent evaluation | Responsible AI | Multilingual NLP
I build and evaluate agentic systems, with a focus on coding-agent controls and trace-level evaluation. My wider research covers responsible AI, multilingual NLP, low-resource African languages, and culturally grounded emotion modelling.
Selected work
Source code, release records, benchmark reports, and research notes are linked with each project. Reported figures come from held-out evaluations or versioned release records.
Research programme
I study what an agent did, how a monitor can identify the point of failure, and which controls should sit before tool execution.
Agent security | Version 0.5.0
An authenticated gateway that evaluates agent tool requests before adapter execution. It checks identity, role, parameters, destination, approval, and session state, then writes redacted hash-linked audit records.
Version 0.5.0 adds requirement-driven causal and temporal probes, bounded policy checking, policy-change analysis, and mutation-tested trajectory contracts.
Monitor benchmark | Version 0.4.1
A contrastive benchmark for testing whether coding-agent monitors ground control decisions in policy-relevant trace evidence. The release contains 320 core traces and a separate 160-trace adversarial holdout.
Evaluation measures unsafe-action prevention, permitted-task retention, evidence grounding, intervention timing, calibration, approval burden, and paired bootstrap intervals.
Python package | Version 0.3.0
African-language text normalisation, selective language routing, code-switch audits, emotion-label mapping, and coverage-aware evaluation. The package runs without an API key or model download.
On 18,402 held-out AfriSenti test tweets, the development-selected rejection policy reached 89.65% accuracy on accepted items at 74.03% coverage.
Dataset | Version 0.1.0
A reproducible dataset of 120 synthetic coding-agent traces across 12 safety and security failure modes, including integrity problems. Each trace links the failure evidence to its first intervention point and a safer counterfactual.
Evaluation harness
A Python harness for scanning agent traces and evaluating deterministic monitors. Findings include localised evidence and a remediation field, with JSON reports for regression testing or later model comparison.
Multilingual benchmark
CPU-friendly valence-arousal baselines for German, English, Nigerian Pidgin, Swahili, and Chinese. Evaluation uses fixed training budgets and grouped resampling, with source-record bootstrap intervals.
Aspect conditioning reduced test macro RMSE from 1.343 to 1.323. The language-adaptive result is reported separately by language.
Open source
These fixes were reviewed and merged by maintainers of established Python AI projects. The pull request pages retain the review discussion and CI record.
Fixed a case where a valid JSON object containing a string produced one query per character. The change added type validation and regression coverage; its release note records the behaviour.
View merged pull requestAdded a one-time warning for labels outside the documented binary form, while preserving existing weighted-loss workflows. Focused tests cover ordinary and non-binary inputs without downloading model weights.
View merged pull requestCitable work
Published records and versioned technical outputs with permanent public links.
Peer-reviewed paper
Proceedings of the 18th International Workshop on Semantic Evaluation, pages 123-132.
ACL Anthology recordConference abstract and presentation
Co-authored work from the DAIL-ICH project, presented at Digital Humanities 2025 in Lisbon. The full title and author record appear in the published book of abstracts.
DH2025 Book of AbstractsSoftware records
Background
I am completing a PhD in Data Science and Artificial Intelligence at the University of Hull. My research examines multilingual emotion classification in code-mixed Afrobeats lyrics, with work on retrieval systems and label provenance.
My recent engineering work covers agent tool policy, coding-agent monitor evaluation, research automation, and reproducible NLP systems.
Professional standing
My community work includes academic support with IntoUniversity Hull East, research mentoring through the University of Hull and Nuffield Research Placements, Tech Help Group, and the Hull Afro-Caribbean Association.
Contact
For research collaboration, technical discussion, speaking, or professional enquiries, use the links here.