Python Engineering for AI, Machine Learning and Data Platforms

Python from notebook to production

Python is the main language for AI and data work. We use it to build AI agents, machine learning pipelines and data platforms on Google Cloud, and we take it past notebooks and prototypes into production with automated testing, type checking and CI/CD.

Where we use Python

AI agents and Vertex AI

We build multi agent systems with the Google Agent Development Kit (ADK), LangGraph, Pydantic AI and Gemini models:

• Structured outputs: Pydantic models enforce JSON schemas so LLM outputs and tool calls are predictable.
• Agent evaluation: LLM as a judge, deterministic assertions and latency profiling run before every release to catch hallucinations and regressions.
• Retrieval augmented generation (RAG): Semantic search with BigQuery Vector Search and Vertex AI embeddings.
• Prompt optimisation: DSPy pipelines to tune prompts against evaluation sets.

FastAPI services on Cloud Run

For microservices and model inference APIs, we build asynchronous FastAPI backends on Cloud Run:

• Async I/O: Non blocking request handling so each container serves more traffic.
• Standards: OpenAPI 3.1 documentation and RFC 9457 Problem Details error responses.
• Lean containers: Multi stage Docker builds to keep images small and cold starts short.

Data engineering on Google Cloud

We build ELT and ETL pipelines for enterprise data:

• PySpark on Serverless Dataproc: Distributed transformations without managing clusters.
• Polars and PyArrow: Fast in memory data processing where Pandas is too slow.
• Dataflow and Apache Beam: Streaming pipelines into BigQuery.
• BigQuery Storage Write API: High throughput streaming ingestion.
• dbt: Tested, version controlled SQL transformations.

Engineering standards

1. Type safety: Type annotations checked with mypy and pyright in strict mode.
2. Dependency management: Locked, reproducible environments with uv or poetry.
3. Testing: pytest suites with fixtures and mocked APIs, run in CI on every change.
4. Security scanning: pip-audit, Bandit and Artifact Analysis container scanning for known vulnerabilities.

Further reading

• MCP and agentic AI on Google Cloud Run: deploying Model Context Protocol servers and agents with Python on Cloud Run.
• The 4 layer architecture of production AI systems: a blueprint for governed AI platforms.
• AI document processing with Vertex AI: document parsing, extraction and compliance scoring with Python pipelines.

Book an engineering architecture review