Contents · September 2026 edition

Research

Rigorous ML research: patient-level validation, uncertainty and public benchmarks.

Showing 2 of 7 entries

Data: 1,027 acquisitions, 627 patients
Patient-level five-fold cross-validation
Architectures and ensembles
Uncertainty and correlated errors
Retrospective triage simulation
Distilled single model
Method outline; thesis figures to follow
01 · Research · 2025 to 2026
MIQA: geometric distortion in prostate DWI
Automated detection of geometric distortion in prostate diffusion-weighted MRI, and an investigation into why the most accurate model was the worst at recognising its own mistakes.
ResultA three-seed DenseNet-121 ensemble reached mean AUROC 0.819 (SD 0.049 across five patient-level folds); a heterogeneous nine-model ensemble kept 83.6% accuracy on the 70% of cases it retained in a retrospective triage simulation.
Read the case study: MIQA: geometric distortion in prostate DWI
- Research evidence
Role
MSc thesis researcher
Status
Retrospectively validated proof of concept
Stack
PyTorch · DenseNet-121 · ResNet-18/50 · EfficientNet-B0 · Deep ensembles
Inspect
Code on GitHub, MIT licence (external link)
02 · Products · 2024 to present
NeuralTake
The AI consultancy I founded in London in 2024, through which I have signed and delivered more than £5,000 of B2B AI software and consultancy work and shipped a production SaaS used by about 50 business users.
ResultMore than £5,000 in signed B2B contracts, delivered; a production SaaS (AI Zuzi) with about 50 business users; a client platform taken from prototype to production with a zero-data-loss live migration.
Read the case study: NeuralTake
- Production deployment
- Commercial validation
Role
Founder and AI engineer
Status
Active; figures are signed contracts, not revenue
Stack
Django · React · PostgreSQL · Stripe · Seven platform integrations
Inspect
neuraltake.com (external link)
03 · Products
NeuralKite
A live product that maps private companies in fragmented UK sectors, scores them against an acquisition thesis and produces decision-ready dossiers; built for people who buy companies, and relevant to AI, private-equity and finance-oriented employers.
ResultDeployed and publicly accessible at neuralkite.com, with sign-in and demo booking; customer and usage figures are not published.
Read the case study: NeuralKite
- Production deployment
Role
Product design and engineering
Status
Live product; customer and usage figures not published
Stack
Next.js · TypeScript · Prisma · PostgreSQL · NextAuth
Inspect
neuralkite.com (external link)
Sparse fridge or pantry input
Retrieval and prompt logic
Multi-step tool flows
Iteration from usage and feedback
Screenshots pending verification
04 · Products
ChefBot
A GPT-based assistant that helps people make meals from sparse fridge or pantry inputs; I designed the retrieval, prompt logic and multi-step tool flows and iterated the experience from usage patterns and user feedback.
ResultLive link and usage figures pending verification; no metrics are published until they can be checked.
Read the case study: ChefBot
Evidence: pending verification
Role
Designer of retrieval, prompt logic and tool flows
Status
Live link and usage figures pending verification
Stack
GPT (OpenAI GPTs) · Retrieval · Prompt logic · Multi-step tool flows
Radiology reports for every study
Report-derived labels
Image model: slices to series
Image-only inference
Weak-supervision pipeline; figures once available
05 · Research · 2026
RSNA Knee Abnormality Detection
An ongoing entry to the 2026 RSNA Knee Abnormality Detection competition on Kaggle: multimodal knee-MRI classification in which only 58 of 4,407 training studies carry gold labels, so free-text radiology reports have to do the labelling.
ResultOngoing as of 15 September 2026, with about a month to go; no medal, leaderboard position or result to report.
Read the case study: RSNA Knee Abnormality Detection
- Public benchmark
Role
Competitor: data engineering, modelling and training
Status
Ongoing; no result yet
Stack
PyTorch · timm · BiGRU · Attention pooling · DistributedDataParallel
Inspect
Competition on Kaggle (external link)
06 · Experiments · 2026
Cash Nova
A fictional producer I created to explore generative music as a creative medium: six albums and 52 tracks on ElevenMusic in 2026, with platform discovery and playlist placement for the work.
Result52 tracks across six albums and 97 followers on ElevenMusic (15 September 2026); tracks received platform discovery and playlist placement.
Read the case study: Cash Nova
- Creative experiment
- Audience reach
Role
Creator and producer
Status
Active
Stack
ElevenMusic · Generative music · Artist identity · Album artwork
Inspect
Listen on ElevenMusic (external link)
07 · Experiments · 2026
CleverFacts and Ogarnik
The same short-form educational format adapted for English and Polish audiences with translation and expressive dubbing, and what the asymmetric results say about which concepts travel between markets.
ResultNearly 720,000 combined views across the English and Polish channels, with approximately 1,000 subscribers on each (15 September 2026); views, not unique viewers.
Read the case study: CleverFacts and Ogarnik
- Audience reach
Role
Creator: format, adaptation and distribution
Status
Active
Stack
Translation · Expressive dubbing · YouTube Shorts · Format testing
Inspect
CleverFacts on YouTube (external link)

MIQA: geometric distortion in prostate DWI

NeuralTake

NeuralKite

ChefBot

RSNA Knee Abnormality Detection

Cash Nova

CleverFacts and Ogarnik