A production platform slice for one agent

A complete Python build of the capstone on synthetic data. A Temporal workflow fed by an idempotent Postgres inbox waits for the lender and the solicitor, runs model-written apportionment code in a locked-down Docker sandbox and recomputes it to the penny, keeps state in Postgres with forced row-level security, brokers each firm's lender key through a credential proxy, routes every model call through LiteLLM with per-firm keys and budgets, and ships prompt changes through a 40-case eval gate in CI.

Back to the project

The slice is built around one workflow per matter, completion-<matter id>, and one rule: the model proposes, code disposes. The model writes the apportionment code and the covering letter; everything that decides anything (whether the code's output is right, which firm a request belongs to, whether a send has already happened, whether a budget allows another call) is code you can read. The decision that matters most is the independent recompute: the sandbox only keeps model-written code from doing harm, and it's validate() recomputing every line in integer pence that stops it from being wrong.

Back to the project

A production platform slice for one agent

A production platform slice for one agent

Sign in to keep reading the answer.

Sign in to keep reading the answer.