Builds the features that call a hosted language model, grounds them in our data, and measures whether the answers hold up.
Build your own roadmap from your CVThe whole application layer, from prompt assembly to the API route, is Python.
The product is built on a model you did not train; wiring prompts, tools and outputs into a working feature is the daily job.
Every feature calls a hosted model through a chat-completions style API; you must know its request shape, its limits and its bill.
The prompt is code you cannot unit test the usual way; you write it, version it and evaluate it like code.
Retrieval, deduplication and search all start with turning text into vectors; you must know what they capture and what they miss.
Most answers come from our documents, not the model's weights; retrieval is how an answer gets grounded.
Every model call sits behind a FastAPI route with a typed request and response; you write and test that route.
A demo that works once is not a feature; you build the eval set and the scoring that says whether a prompt change helped.
Grounding means finding the right chunk fast; you index, query and tune a vector store.
Prompts, evals and routes ship through pull requests; a prompt nobody can diff is a prompt nobody can fix.
The model acts through tools you define; you write the schema, the handler and the guard around it.
Every model output is parsed into a typed schema before anything trusts it; Pydantic is that gate.
The service ships as a container; you build it, run it locally and make it start the same way in staging.
User text reaches the model and the model's text reaches users; you put the checks on both sides.