Most of these were built inside client engagements, so the source isn't public.
Where a repository or demo exists, it's linked on the card.
Aerial Imagery Inference Pipeline
Production
High-throughput PyTorch inference for aerial imagery, used to assess property risk and process insurance claims.
An end-to-end deployment and development pipeline for aerial computer-vision models. The work centred on making inference fast and cheap enough to run at portfolio scale: GPU-accelerated batching, memory-efficient tensor operations, and ONNX export to shrink the runtime footprint of the deployed models.
Problem
Insurance risk assessment and claims triage needed image analysis across a large and continuously growing volume of aerial imagery. Naive per-image PyTorch inference was too slow and too memory-hungry to run economically at that volume.
Approach
Rebuilt the inference path around GPU-accelerated batched execution and memory-efficient tensor operations, and moved deployed models to ONNX to reduce their memory footprint. Wrapped the result in scalable serving infrastructure so throughput scales with demand rather than headcount.
- Python
- PyTorch
- ONNX
- CUDA
- AWS
- Docker
Enterprise RAG & Vector Search Platform
Production
End-to-end retrieval-augmented generation pipeline — ingestion, embedding generation, vector search and inference — for enterprise applications.
A complete Python pipeline taking enterprise documents from raw ingestion through embedding generation into a vector index, then serving retrieval and generation for downstream applications.
Problem
Enterprise teams needed to query large internal document corpora in natural language, with answers grounded in their own content rather than in a general-purpose model's training data.
Approach
Designed and implemented the full pipeline in Python: ingestion and normalisation, embedding generation, vector storage and search, and the inference layer that assembles retrieved context into grounded responses.
- Python
- RAG
- Vector Search
- Embeddings
- FastAPI
- AWS
Claim Notes Feature Repository
Production
A data pipeline that derives structured flags from free-text claim notes, acting as the central feature store for predictive models.
Unstructured adjuster notes contain most of the early signal about a claim, but none of it in a form a model can consume. This pipeline extracts that signal into structured flags and centralises them so every downstream model reads the same feature definitions.
Problem
Predictive models across the claims organisation each re-derived features from raw notes, producing inconsistent definitions and duplicated work.
Approach
Built a pipeline, designed with the data-analyst team, that generates flags from claim notes and publishes them to a single central repository that all predictive models consume as their feature source.
- Python
- Spark SQL
- Airflow
- AWS
- NLP
Analytical Ecosystem Framework
Delivered
A PySpark framework that fuses first-party and third-party data sources into model-ready datasets and insights.
A reusable ingestion and transformation framework rather than a one-off pipeline — new data sources plug into the same processing path and land in the same monthly-refreshed schema.
Problem
First-party and third-party data arrived in different shapes, on different cadences, and had to be reconciled by hand before any modelling could start.
Approach
Designed and developed the framework pipeline in PySpark, with Spark and Hive ETL underneath and Python jobs loading the unified schema on a monthly interval. Tableau dashboards automated the quality-check process.
- PySpark
- Apache Spark
- Hive
- Python
- Tableau
Automated Test-Failure Triage Tool
Delivered
Python tooling that captures product test failures and files bugs automatically, backed by an auto-generated regex per failure signature.
Triaging test failures by hand consumed engineering time that scaled with the test suite. This tool closed the loop between a failing test and a tracked bug without a human in the middle.
Problem
Product test failures were captured and filed manually, which was slow and produced inconsistent, hard-to-deduplicate bug reports.
Approach
Built a Python tool that detects failures, generates a regex describing each failure signature, and opens a bug backed by that signature — making repeat failures automatically recognisable.
- Python
- Regex
- Automation
- CI