BDPE: production AI in under thirty days.
What we shipped.
The first BDPE workflow takes the firm's source documents and produces a structured first-pass analysis in the firm's house style. It runs on every matter that fits the workflow's scope, replacing a manual step that previously sat with associates.
The detailed input/output specification — what counts as a valid input, what the output schema looks like, what the firm considers a correct answer — is captured in the test set the workflow is held to.
What the firm supplied.
BDPE's domain experts supplied a body of completed work — not prompts, not specifications, just inputs and outputs from real matters they had handled correctly. Edge cases were captured the same way: examples of unusual inputs and the output an expert produced for them.
The example set was the contract. Podesta's job was to produce a workflow that, given a held-out input from the same set, produced an output an expert would accept.
Examples to tests to workflow.
Podesta converts an example set into a test set automatically: each example becomes a test case the workflow has to pass. The workflow itself is then generated and iterated until every test passes.
Every iteration was reviewed by a BDPE expert before publication. When a test failed, the failure was visible to the expert with the exact input, the expected output, and the workflow's actual output. Where the expert disagreed with the test, the test was changed; where the expert agreed, the workflow had to change.
Under thirty days, end to end.
Workflow scoped jointly with BDPE. First example set delivered.
Tests generated from examples. Workflow built and iterated to passing on the first cohort of cases.
Edge-case examples added; workflow iterated to passing on the full set. Expert review on every change.
Workflow live, called from BDPE's platform via the Podesta SDK.
Called from BDPE's platform.
BDPE's developers integrated the workflow into their internal patent analysis platform using the Podesta SDK. It's called from code like any other typed service — there is no AI logic in the firm's repo, no prompts to maintain, no model upgrades that require a release.
When BDPE updates the workflow, they update it in Podesta — not in their codebase. Domain experts publish changes after they pass tests; engineering doesn't ship a release for an AI behaviour change.