A narrated financial recap video, generated end to end
A generated video is only worth shipping if you can say afterwards which model produced each frame of it, and prove the file has not changed since. Generative media is the easy half; provenance is the half that decides whether the output can be trusted.
Independent system · pipeline, media generation, storage, and delivery
The pipeline ran end to end in roughly two to four minutes, dominated by generating five scene images in parallel, and shipped with a Plaid sandbox connector as a genuine no-credentials path.
Banks hold a year of a customer's transactions and hand it back as a table nobody opens. There is no moment that makes the money mean anything, so the app goes unused and the relationship goes with it.
Scene images, narration audio and the narrative script all leave through a single client wrapper, with a fallback provider behind the same interface. Zero direct provider calls exist outside that file, which is what makes a per-artifact provenance record possible at all.
Fourteen files across ten artifact types — video, thumbnail, script, analytics, prompts, generation provenance, five scenes, narration, the source CSV and metadata — are written to Backblaze B2. SQLite is a read cache, so a redeploy loses nothing, and a 45-day lifecycle rule plus a per-artifact SHA-256 makes retention and integrity checkable.
Each scene renders to its own segment and the segments are concat-joined with the narration, dipping to black between them. Rendering the whole timeline in one pass is simpler and needs more memory than a hobby-tier host has, so the compositor was rebuilt to stay bounded and off the event loop.
The hosted frontend and API were taken down after the hackathon; both URLs now return 404, and the project did not place. The source, the architecture records and the recorded demo remain as the evidence.
I’m always interested in the constraints that make an architecture worth explaining.