Back to portfolio

Python Microservices for LLM Data Import and Inference (2023)

A three-week senior backend engagement in July 2023: Python microservices for importing and transforming data for large language models, inference functions, test suites, and deployment on GCP managed services with Docker.

Architecture: source data, import service, transformation, then inference functions, GCP managed services and Docker deployment
pythonmicroservicesllm data pipelinesgcpdocker

The Brief

A product team working to tight deadlines needed a senior Python backend developer to speed them up. Their system was a set of microservices whose job was to import data from several sources, transform it into the shape a large language model could use, and run inference functions over it. The brief was explicit about what mattered: well-structured code, non-functional requirements met, comprehensive tests, and services that stayed available and scalable on Google Cloud's managed services.

What I Did

  • Import and transformation microservices: new functionality in the Python services that pull in records, validate and normalise them, and turn them into the text and structure the model consumes.
  • Inference functions: the services that call the model and hand back structured results to the rest of the platform.
  • Tests as part of the deliverable: unit, integration and component tests written alongside the code, because the team's deadline left no room for regressions found later.
  • Working inside the team's architecture: Docker images and GCP managed services the DevOps and infrastructure teams already ran, so nothing I built needed a new operational path.
Jul 2023three weeks, hourly, closed and reviewed
2kinds of service: data import and transformation, inference
3test levels shipped with the code: unit, integration, component
GCPmanaged services and Docker, the team's existing platform

Why It Holds Up

Most LLM work in 2023 was prompts. The unglamorous part, getting messy source data into a clean, tested pipeline the model can rely on, is where production systems succeed or fail, and it is the part this engagement was about. It is also why I approach every AI feature from the data side first: if the import is wrong, no prompt fixes it.

Message me on the platform where you found this and I will answer there.