Known By Their Actions: Fingerprinting LLM Browser Agents via UI Traces
NeurIPS 2026
UI interaction traces can fingerprint the model behind a browser agent with up to 96% F1.
I am a PhD researcher at the University of Oxford, advised by Scott A. Hale. I work on factuality of LLMs and on training methods that ground model outputs in evidence. My research is funded by the Rhodes Scholarship.
I recently spent time as a Student Researcher at Google DeepMind, working with Miruna Pîslar on agentic student simulators for advancing AI tutors. Before that, I completed two AI Research Internships at JPMorganChase AI Research Lab, where I developed a synthetic data generation framework for financial NLI and a post-training recipe for a specialised 8B claim decomposition model.
I've worked on:
- verifying the factuality of language model outputs (claim decomposition, robustness of embedding models)
- identifying agents from behavioural traces (browser-agent fingerprinting)
- testing whether benchmarks measure what they claim to measure (construct validity, linguistic reasoning and memorization)
- frameworks for evaluating bias in T2I systems (community-centred evaluation, participatory data work)
Previously, I worked as a Software Engineer for four years at Twiga Foods, a startup solving food security challenges across Africa, where I built Kotlin
and Go
microservices and Android
application features to support over 40,000 farmers and vendors.
News
Recent updates
- Known By Their Actions was accepted to NeurIPS 2026.
- INCLUDE-2.0 was accepted to EMNLP 2026 Main as an oral.
- Distill and Align Decomposition (DAD) was accepted to Findings of EACL 2026.
- Started as a Student Researcher at Google DeepMind.
- Measuring what Matters was accepted to NeurIPS 2025 Datasets & Benchmarks.
- Co-leading the LLM practical at Deep Learning Indaba Kigali 2025.
- World Wide Recipe received a Best Paper Honourable Mention at FAccT 2025.
- Started an AI Research Internship at JPMorganChase AI Research (London).
Publications
An asterisk marks equal contribution.
2026
INCLUDE-2.0: Disentangling Knowledge Gaps in Multilingual LLM Evaluation
EMNLP 2026 Main Oral
The benchmark separates language proficiency from regional and community-grounded knowledge across 91 languages.
Distill and Align Decomposition for Enhanced Claim Verification
EACL 2026 (Findings)
A distilled, reinforcement-learned 8B decomposer improves verification while preserving the quality of complex subclaims.
2025
Measuring what Matters: Construct Validity in Large Language Model Benchmarks
NeurIPS 2025, Datasets & Benchmarks
A review of 445 benchmarks identifies recurring validity gaps and gives eight recommendations for better LLM evaluation.
When Claims Evolve: Evaluating and Enhancing the Robustness of Embedding Models Against Misinformation Edits
ACL 2025 (Findings)
Realistic edits make claim retrieval brittle, but targeted mitigation substantially improves in- and out-of-domain robustness.
The World Wide Recipe: A Community-Centred Framework for Fine-Grained Data Collection and Regional Bias Operationalisation
FAccT 2025 🏅 Best Paper Honourable Mention
Community-led data collection exposes regional inaccuracies and cultural misrepresentation in text-to-image systems.
FinNLI: Novel Dataset for Multi-Genre Financial Natural Language Inference Benchmarking
Findings of NAACL 2025
A 21,304-pair benchmark shows substantial domain-shift failures in financial natural language inference.
A Variational Approach for Mitigating Entity Bias in Relation Extraction
ACL 2025 Main
A variational information bottleneck reduces reliance on entity-specific shortcuts while retaining task-relevant information.
The Human Labour of Data Work: Capturing Cultural Diversity through World Wide Dishes
CSCW 2025
A reflexive account of the trust-building, access work, and cultural mediation required to build participatory datasets.
2024
LINGOLY: A Benchmark of Olympiad-Level Linguistic Reasoning Puzzles in Low-Resource and Extinct Languages
NeurIPS 2024, Datasets & Benchmarks Oral · Top 0.5%
Models struggle with difficult multi-step linguistic reasoning across more than 90 mostly low-resource languages.
Scaling Crowdsourced Election Monitoring: Construction and Evaluation of Classification Models for Multilingual and Cross-Domain Classification Settings
AAAI Workshop on AI for Credible Elections 2024
Multilingual classifiers help triage crowdsourced election reports, while exposing performance gaps between English and Swahili.
Projects
Visual Diagnoser
2019An Android application for home-based diabetic retinopathy screening. We trained a CNN in TensorFlow to classify retinal images with diabetic retinopathy and deployed it to low-end mobile devices using TensorFlow Lite.
Ms Maji
2018An IoT-based water management system designed for densely populated communities, built using Arduino and the IBM Watson IoT platform.
Elimu
2019A personalised learning application using Deep Knowledge Tracing with LSTM networks to model student knowledge and adapt learning activities over time.
Teaching
Tutorials and talks
Tutorials
Developer talks
