Teku AI builds a parcel-inspection app with Hermes and Jev’s experimental image support. The walkthrough’s strongest evidence is an application bottleneck: waiting for each image’s response before sending the next.

01:49–09:14: separate the builder from the classifier

  • Hermes builds and deploys the app; Jev receives images and answers predefined inspection questions.
  • The accompanying build guide specifies server-side credentials, authentication and request limits.
  • Its integration asks six yes/no questions: box presence, overall damage and four damage categories.
  • Local code chooses the verdict from the overall damage answer; damage-category scores supply detail.

13:08–16:02: measure the whole batch separately

  • Teku reports roughly 10 seconds for 20 images in the first deployed batch and about 3 seconds after fixing sequential submission.
  • Those are presenter observations from one demonstration; neither the timings nor the model comparison were independently reproduced.
  • The guide specifies a bounded worker pool, so uploads overlap without sending an unlimited batch at once.
  • It separates server-measured API time from the browser’s first-upload-to-last-response stopwatch.

“Photos are analyzed in parallel, never one at a time.” — Teku AI’s accompanying build guide.

Review flags and errors need their own paths

  • The guide adds a review flag around a chosen damage threshold while keeping the intact/damaged verdict visible.
  • Its thresholds are example policy settings. Their scores need validation on held-out images before being treated as calibrated risk.
  • API errors remain errors, rather than becoming an “intact” verdict; the guide calls for bounded retries and per-image failure displays.
  • A typed answer and a polished inspection animation provide no guarantee that the parcel was classified correctly.

The headline is a projection, not a thousand-image test

The title extrapolates cost from a small batch. Narration, description and guide disagree on some timing and price figures, so they cannot support a precise general cost or speed claim. The guide also reports improvements on reused stock photos and explicitly calls for fresh validation.

Caveat: the full transcript and linked guide were reviewed; the generated application was neither inspected as source code nor run here. Camera monitoring and the proposed business deployments are follow-up ideas.