Submission + - MicroMachina: Turbo Music Video Generator (Shyamal Chandra of Crankshaft News) (github.io)
advancecoder writes: Programmatic short-form advertising places demanding requirements on a generative pipeline: hyper-dense dialogue that remains intelligible, character consistency across independently generated clips, bounded cost, and resilience to the noisy, non-deterministic behavior of hosted generative-AI APIs. We present MicroMachina, a production web application that synthesizes sixty-second “fast-talk” commercial videos from a single textual pitch. The system couples an LLM-driven script director with a four-stage asset pipeline—text-to-speech, text-to-video, music, and ffmpeg muxing—and adds three properties that are essential for industrial use: resumability (a disk-backed checkpoint manifest that turns any failure into a resume point), resilience (bounded retries with backoff and semantic retry classification), and financial safety (a confirmation gate before spend and a user-invokable hard cancel that propagates abort signals into in-flight requests). We report the system architecture, the failure taxonomy, and the energy (consumption-budget) model. We validate the reliability-critical components with a hermetic test suite that exercises retry, cancel, and resume logic without incurring API cost.
