THE ROLE
You will build the systems behind our internal AI intelligence layer and our client AI work. This is applied engineering on production systems with real users inside the company and real clients outside it. It is not a research post, and it is not a prompt writing post.
The work is varied by design. In a given quarter you might tune a retrieval pipeline that is returning the wrong chunks, design an authorisation model so an agent cannot read what its caller cannot read, benchmark two open weight models on a client's data to justify an on premises deployment, and then write the evaluation harness that proves the change was actually an improvement.
WHAT YOU WILL DO
Build and maintain retrieval pipelines, covering chunking strategy, embedding selection, vector store performance and hybrid search behaviour. We run pgvector on PostgreSQL today.
Design and ship agent workflows and tool calling systems, with attention to failure handling, cost and latency rather than demo quality output.
Own model evaluation. Build the harnesses,
define the metrics and produce the evidence that a change helped. We would rather have an unglamorous eval suite than a confident opinion.
Deploy and operate inference workloads on GCP and Vertex AI, and contribute to sizing and configuration for on premises open weight model hosting.
Integrate AI capability into the wider product surface through APIs, webhooks and event driven pipelines.
Work with the AI Product Owner to turn requirements into technical designs, and push back when the requirement and the technology disagree.
WHAT WE ARE LOOKING FOR
Strong Python, and comfort in a production codebase rather than only in notebooks.
Practical experience with LLM applications: RAG systems, agent frameworks, structured output, function calling. You have debugged a retrieval system that was quietly returning rubbish, and you know how you found it.
Solid SQL and PostgreSQL. Vector database experience is a transparent advantage.
Cloud deployment exper