Senior Full-Stack Engineer and Data Scientist based in the United Kingdom, building AI systems, backend platforms, and real-time computer vision solutions.
I take end-to-end ownership from system architecture and model integration through APIs, data pipelines, and production delivery, with a focus on computer vision, hyperspectral imaging, and edge inference.
Open to senior opportunities in applied AI, computer vision, ML systems, and backend engineering.
Screening preview with experimental statistical models. Investigating road collision evidence alongside traffic exposure, rather than ranking roads by collision counts alone.
- Implemented: reproducible DfT data acquisition, collision-to-road matching, exposure joins, FastAPI endpoints, and a React/MapLibre investigation interface.
- Modelling: an exposure-rate baseline, experimental negative-binomial Safety Performance Function, and Poisson-gamma Empirical Bayes retrospective screening.
- Boundary: geographic holdout evaluation still needs unseen-authority evidence. These are screening and research tools, not validated future-risk predictions or intervention recommendations.
Methodology | Architecture | Run locally
Private, in development. A computer-vision project for second-hand fashion intake, image-quality checks, and evidence-linked listings. The focus is traceable visual evidence and model evaluation; this is not yet a public end-to-end product demo.
Early-stage spatial-data foundation. Building towards explainable urban heat analysis for Greater London using Earth-observation and environmental data.
- Implemented: dataset registry, study-area contract, catalogue audit, and STAC scene discovery for Sentinel-2 and Landsat.
- Next: validated spatial acquisition and model evaluation. Heat prediction, intervention scenarios, and optional airborne hyperspectral analysis remain planned work, not demonstrated results.
Dataset feasibility | Architecture | Roadmap
Upstream work in computer vision, geometry, model validation, and backend reliability. Each link points to the implementation and review history.
| Project | Contribution |
|---|---|
| Assimp #6872 | Accept integer texture UV transform components in FBX imports. |
| Kornia #4355 | Handle z1 unprojection depth according to tensor rank. |
| PCL #6457 | Extend BruteForce search to non-XYZ point types. |
| Open3D #7525 | Correct Gaussian-splat Vulkan render-target setup. |
| OpenCV #29622 | Accept CV_Bool masks in connected-components processing. |
| ONNX #7821 | Validate grouped input-channel constraints for ConvTranspose. |
| MLflow #22653 | Add a configurable Huey storage URL for Redis-backed server jobs. |
| Immich #28884 | Skip existing album users without failing mixed sharing requests. |
More merged upstream PRs | Open upstream PRs
- Kapdaa: built an AI textile-sorting stack spanning backend services, APIs, UI, computer vision, hyperspectral models, and edge deployment. Reported results include
>90%garment-type classification across15+classes, fibre-classificationF1 > 0.90, and30-40 garments/minthroughput. - Jeavio: delivered enterprise backend features and distributed-system improvements, including targeted Go rewrites with approximately
30%execution-efficiency improvement.
These are professional-work results, separate from the public research projects above.
- Languages: Python, C++, Go, TypeScript/JavaScript, SQL, Bash
- AI and computer vision: PyTorch, TensorFlow, OpenCV, Hugging Face, vision transformers, hyperspectral imaging, ONNX
- Edge inference: NVIDIA Jetson, TensorRT, DeepStream, OpenVINO
- Backend and data: FastAPI, Flask, Django, REST, GraphQL, gRPC, PostgreSQL, Redis, Kafka, Spark
- Delivery: Docker, Kubernetes, CI/CD, AWS, Azure, GCP
Last updated: 30 September 2026. Project descriptions reflect the published repository documentation; research goals are distinguished from implemented capabilities.