Research
Open questions, current answers
Four questions organise the current programme. Each entry below states what it is trying to establish, so an unfinished result is still worth reading.
Verifiable generation
How a generated statement can be tied to evidence strongly enough that an ungrounded claim is detectable automatically rather than by reading. Most of my applied work depends on this being solvable.
Speech models at the edge
What accuracy actually costs when a recognition model has to run on consumer hardware, and which parts of a speech pipeline degrade gracefully under quantisation and which fail outright.
Evaluation under distribution shift
Why benchmark performance predicts production performance so poorly, and how to build test sets that fail early on the inputs a system will actually meet.
Anomaly detection with generative models
Using learned representations of normal behaviour to find fraud and misuse, including where a generative approach outperforms a discriminative one and where it quietly does not.
In progress
Current work
- Technical report2025In progress
Grounding Clinical Note Generation Against Transcript Evidence
How reliably can span level grounding detect fabricated clinical content before a clinician sees it?
A working report on the evaluation approach behind Nora Health. Summarisation metrics reward fluency and are close to blind to fabrication, which is the failure that matters in a clinical record. This report sets out an evaluation that decomposes a generated note into individual clinical assertions, requires each to anchor to a transcript span, and reports grounding separately from readability. It includes the error taxonomy that came out of structured review sessions with clinicians.
- BERTScore
- Span alignment
- Error taxonomy
- Human review protocol
- Experiment2025In progress
Quantisation Trade Offs in On Device Speech Recognition
Where does word error rate break under quantisation, and is the degradation uniform across clinical vocabulary?
A controlled comparison of quantisation strategies for speech recognition models running on consumer hardware, measuring word error rate against latency and memory. The finding that shaped the product is that degradation is not uniform: general conversational accuracy holds up considerably better than drug names, dosages, and abbreviations, which is exactly the vocabulary a clinical system cannot afford to get wrong.
- Word error rate analysis
- Post training quantisation
- ONNX Runtime
- Latency profiling
Complete
Publications and prior research
- Thesis2023Complete
Generative AI for Fraud Detection in PropTech
Can adversarially trained generative models detect fraudulent property transactions that rule based and supervised systems miss?
An investigation into generative adversarial approaches for anomaly detection in real estate transaction monitoring. Property fraud is rare, adversarial, and constantly changing shape, which is precisely the setting where supervised classifiers trained on historical labels perform worst. The work trains a model of legitimate transaction structure and treats reconstruction difficulty as the fraud signal, then examines where that assumption holds and where an adversary can exploit it.
- Generative adversarial networks
- Anomaly detection
- PyTorch
- Imbalanced learning
- Research project2022Complete
Hybrid PCA and Factor Analysis for Clinical Predictive Models
Does combining principal component analysis with factor analysis improve stability in healthcare prediction without losing interpretability?
Work on feature reduction for medical predictive modelling, where the number of correlated measurements is high, sample sizes are limited, and a model nobody can interpret will not be used regardless of its accuracy. The project compared a hybrid reduction approach against each method alone, measuring both predictive stability across resamples and how well the resulting components mapped to clinically meaningful structure.
- Principal component analysis
- Factor analysis
- Cross validation
- Statistical modelling
4 entries. Technical reports are published here rather than submitted, so they are dated and revised in place as the work changes.
Contact
If you are working on something where being wrong matters, I would like to hear about it.
I am open to consulting engagements, research collaborations, and conversations that do not have a clear outcome yet.