A model decision carried from experiment into two product builds.
A recorded speech-recognition comparison used a Mac and DGX Spark as complementary tools, then carried the selected approach into UtterFlow and AirSticky with build verification.
04 / proof surface
01 / The challenge
Start with the operating tension.
Model and runtime choices only matter when they improve the actual product under real constraints. The evaluation needed to connect quality and resource tradeoffs to a bounded implementation decision.
02 / The approach
Shape the system around the constraint.
The work used a structured comparison, preserved the finding and its limits, then applied the chosen transcription change to two spatial products. The compute split was treated as a placement decision, not a claim that heavier hardware is always better.
03 / The working system
What exists.
- 01
Recorded ASR and transcription-cleanup comparison
- 02
Mac and DGX Spark roles chosen by task and runtime needs
- 03
A bounded implementation change across two spatial apps
- 04
Build verification recorded after the change
Capabilities demonstrated
Transferable judgment, shown in context.
04 / What the evidence supports
Proof on the page.
- The reviewed record contains a 368-session comparison and a defined finding.
- UtterFlow and AirSticky build verification was recorded after the change was carried across.
- The multi-device history distinguishes the current single-Spark posture from a retired two-node experiment.
Claim boundary
Where the evidence stops.
- The comparison must be reproduced with public-safe fixtures before full technical publication.
- No universal model-superiority, latency, cost, or production-reliability claim is made.
- A successful build is not the same as measured user impact.
Explore the artifact
Next case
05 / Field product · domain translationFieldStack