Chris – Python, Tensorflow, PostgreSQL

A leading provider of AI powered vocational education assessments. Chris was employed to work on a grant funded project to research, build and deploy video-based assessments. The use cases included maintenance engineer and hospitality & tourism qualifications.

• Designed a multi-modal assessment service powered by Gemini 3.1 Pro on Vertex AI, paired with a Langfuse-driven evaluation pipeline, running serverlessly on Cloud Run
• Integrated Gemini 3.1 via Vertex AI for video ingestion, evidence extraction and structured competency scoring.
• Architected the GCP deployment: Cloud Run services, Cloud Build CI, Firestore for assessment and feedback records, and resumable GCS video uploads fronted by Firebase Hosting with least-privilege IAM service accounts throughout.
• Built a Langfuse-based evaluation pipeline covering dataset curation, LLM-judge scoring, regression testing and trace-level performance analytics.

An AI powered procurement and accounts payable platform for multi-site operators in hospitality, retail, construction, healthcare and education. I worked directly with the founder to design and build the backend services, including document processing pipeline, analytics and chat agent.

• Built the document understanding pipeline and supporting backend services which provides an 80% reduction in document (invoice, purchase order, etc) processing time.
• Built LEDGE, the platform's a multi-turn, streaming agent agent, on LangGraph, with authentication passthrough to the platform API, tool calling over the analytics and action endpoints, and a sandboxed Python execution tool for ad-hoc data analysis
• Designed FastAPI / Pydantic services combining LLM-based extractors with OCR for invoices, receipts and statements, evaluated continuously against ledger ground truth
• Authored the Pulumi (Python) program provisioning the AWS estate across dev and prod stacks: Cognito, S3, SES, SQS intake queue + DLQ, and the intake Lambdas, GitHub Actions OIDC deploys, multi-account AWS SSO and Secrets Manager configuration

The company provides the Cambridge English Dictionary (print and online) and several other English language and education products. Chris was hired to explore the use of generative AI to develop new product offerings.

He successfully demonstrated the potential for new product offerings built using AI services by developing several prototypes utilising OpenAI and AzureAI (as well as Streamlit and NiceGUI for demo frontends) including an automated lesson plan generator for English teachers, a Q&A English revision tool and a pronunciation checker. These were later taken in house for production development.

A desktop application that allows sonar analysts to efficiently label large amounts of sonar data for training machine learning (ML) models.

Under the hood, SoLT combines two techniques for streamlining the labeling process:

1. Weak Supervision is an ML technique that involves creating proxy labels from many labelling functions, each of which captures a simple heuristic a domain expert would use to guess the correct label;
2. Active Learning is an iterative approach for selecting examples from the dataset for manual labelling that will have the biggest impact on model performance. The UI was built using Kivy, a Python library for developing cross-platform GUIs.

• Liaised with stakeholders and captured the client's requirements;
• Managed the team and development roadmap;
• Implemented the core algorithms, data pipeline, and Python interface;
• Wrote unit tests using pytest and set up CI/CD process;
• Trained ML models using Tensorflow for demonstration;
• Performed exploratory data analysis and visualisation.

Foresense consumes high-frequency time series data from machine sensors and uses advanced machine learning techniques to detect anomalies and forecast the probability of faults and failures.

• Designed system architecture;
• Designed and built a data pipeline;
• Designed and built deep learning models for anomaly detection and time series forecasting;
• Helped with user research and capturing requirements;
• Managed agile team;
• Developed product roadmap;
• Helped with API design and implementation;
• Helped with containerisation and CI/CD.

SPEaC, aka "The Happy App" was an application (web and mobile) that was designed to increase employee engagement and well-being by providing a way for employees to leave anonymous feedback. The app aggregated feedback to indicate the overall mood of each team/department and enabled managers to respond and create actions and reports based on the feedback. The app was deployed in several NHS Trusts from 2015-2023 and won an innovation award from the British Medical Journal, and was the subject of published research. Unfortunately, a viable business model was not found, and the app has recently been closed down.

• Designed and developed full stack web app (Postgres, Flask, JQuery, Bootstrap);
• Managed deployment on Heroku;
• Implemented multi-tenant architecture;
• Developed API for external integration;
• Managed first-line support;
• Collaborated with external contractors.

Satellites For Batteries (S4B) estimates the probability of battery metal deposits using ensemble machine learning and combining various types of remote sensing and geological data as inputs. It also has a prototype UI built using Streamlit that allows domain experts to visualise model predictions and understand the ML model's behaviour using LIME explainability.

• Developed a modular and reproducible data pipeline using DVC;
• Used GeoPandas and various other Python libraries;
• Performed exploratory data analysis and visualisation;
• Developed an ensemble machine learning model to predict the probability of mineral deposits;
• Designed UI in Streamlit and implemented explainability using LIME
• Led agile data science team;
• Collaborated with stakeholders from multiple companies.