Applied AI / ML

Start with a useful task

Extracting details from a document, answering a question, detecting an unusual transaction, and predicting churn require different data, evaluation criteria, and workflows.

My personal engineering portfolio covers ten areas, including data preparation, modelling, evaluation, APIs, tests, and operational tooling. These implementations help me test architectural ideas and discuss concrete scenarios with clients.

A map of tasks and implementations

Business question	Implementation area	What adoption must validate
Who may leave?	Churn prediction and experiment tracking	Leakage, predictive usefulness, cost of intervention
Where is the answer?	Document retrieval and RAG	Search quality, sources, permissions, refusal behaviour
What does the text contain?	NER and entity spans	Annotation quality, entity classes, document-level errors
Which relationships look suspicious?	Graph analysis and fraud models	Explainability and false-positive cost
What is unusual in this stream?	Streaming anomaly detection	Latency, signal quality, operational response
What type of document is this photo?	Computer vision and inference preparation	Image quality, rare classes, applicable conditions
What is this property worth?	Regression and factor explanation	Representative data and prediction stability
Can review effort be reduced?	LLM code-review workflows and adaptation experiments	False findings, security, verification quality
What should we recommend?	Similarity search and feature management	Relevance, freshness, feedback
Can we trust the incoming data?	Quality checks and distribution monitoring	Materiality and incident handling

The model is part of a larger system

A useful deployment needs input contracts, response formats, latency limits, error handling, and a way to observe quality over time. For example, an NER service must pass the extracted organisation name downstream with its position in the original document. A person needs to be able to correct an error, and the system needs to preserve that correction for quality analysis. A RAG answer must show the material on which it is based.

This transition from model to working process is where my architecture and leadership experience becomes useful.

Background and scope

In 2020 I completed practical Data Science work, from preprocessing to computer vision. My enterprise experience includes ML/AI requirements, data preparation, prototypes, graph algorithms in Mirion, and leading an area that included NER and ML data annotation services.

The case studies describe my enterprise leadership contribution. This page covers personal engineering implementations; it does not imply ten client deployments.

Work with me

We can identify a useful scenario, assess data, define evaluation criteria, and build a prototype. Evidence of value then guides integration, operations, and further development.

Demonstrate one working decision

For documents, follow text through entity extraction to a verifiable API response. For prediction, inspect features, evaluation and the action triggered by a score. For a stream, follow an event into an actionable signal.

A pilot needs an evaluation set and an agreed cost of errors. Then we can decide what to integrate, adapt or replace with a simpler rule. Performance in a personal prototype does not guarantee performance in your domain.

Discuss your AI/ML challenge →