Applied ML for problems that matter
I work with organisations that need machine learning done properly — where reliability, explainability, and domain fit matter as much as benchmark performance. My engagements combine research-grade rigour with practical engineering: systems that work in production, not just in notebooks.
Knowledge extraction & NLP
Extracting structured information from unstructured professional text — legal contracts, technical documents, operational reports. Combining FST-based methods with neural sequence models for high-precision, auditable outputs.
Computer vision systems
Object detection, tracking, and flow analysis for operational environments. Experience deploying real-time CV systems at infrastructure scale, with attention to false-positive management and explainability for operational staff.
ML architecture & integration
Designing ML systems that fit into existing operational workflows — data pipelines, inference serving, monitoring, and evaluation infrastructure. Particular experience with edge deployment and constrained-hardware environments.
Research consulting
Technical advisory for research teams at the intersection of symbolic and neural approaches. Literature review, experimental design, and critical evaluation of neuro-symbolic architectures for specific application domains.
Airport infrastructure computer vision
Real-time object detection and people-flow analytics system deployed across 12 camera zones in Heathrow Terminal 5. TensorRT-accelerated inference on edge hardware, with an operational dashboard for terminal supervisors. Sub-30ms latency at 1080p, false-positive rate below 2% on safety events.
Infrastructure · 2023
Knowledge extraction pipeline extension
Extending an existing NLP pipeline for structured information extraction from commercial contracts. Added FST-based clause detection, argument extraction, and a schema alignment layer that maps raw extractions to a typed knowledge graph. Precision 94%, recall 98.7% on in-distribution relation types.
Legal tech · 2023
Let's work on something hard
If you're working on a problem where standard ML approaches aren't quite right — where you need interpretability, domain fit, or something that actually works in production — I'd like to hear about it.
Get in touch