Case study 01 / Platform architecture
One product across image, video, voice, and music models
How a fragmented catalogue of image, video, speech, music, and editing models became one learnable product with consistent jobs, credits, media, and failure handling.
The main engineering challenge was a stable job, credit, storage, and recovery flow across providers that behave differently.
01 / Challenge
Every model arrived with a different operating model
Providers expose different inputs, output formats, processing times, pricing units, and failure modes. Passing those differences directly to users would have produced a catalogue of disconnected forms rather than a coherent creation platform.
02 / System response
Treat providers as infrastructure, not product surfaces
Piff was designed around stable product concepts: a generation request, an upfront credit quote, a durable job, a media result, and a recoverable failure. Provider-specific settings remain available where they add creative value, but the lifecycle around them stays predictable.
- 01Normalise heterogeneous providers behind one job lifecycle
- 02Quote transparent credit cost before a generation is submitted
- 03Separate orchestration, accounting, storage, and interface feedback
- 04Design explicit retry and failure paths for asynchronous providers
- 05Build reusable character workflows for sequences of up to 15 linked images
03 / Result
A platform users can learn once
The product supports image, video, speech, music, avatars, lip sync, and eight editing tools without forcing users to learn a new workflow for every model. Piff has passed 10,000 completed generations and 1,000 creators, with a published rating of 4.8/5.