Celia.ai

The inversion where deterministic curated data decides and the LLM only explains is non-obvious, and the HarmonyOS-native combination of barcode checks, watch vitals, emergency card and intents is not a port of an existing app. C's 7 fairly notes that medicine-check apps exist, but the evidence supports the members who found the device-level design fresh.

The use case maps to concrete patient moments: barcode checks against about 68k Polish packs, a 13-language emergency card, an SOS chain and an offline safety core. C's 6 overweights the narrow condition scope, which this human-centric brief explicitly rewards, so the median 8 stands.

Measured facts corroborate the claims: 594 test cases across phone, watch and backend, 388 in-event commits by 3 authors, real error handling and honest status labelling. C's 7 rests on tests fixed during the event and count discrepancies, but the measured count is 594 and in-event fixes are normal iteration.

About 20 OHOS kits in real use beyond a webview, including Intents Kit, Form Kit, Live View, Scan and Vision, NFC and a standalone wearable app that would not run unchanged on another OS. A's 8 penalises unverified headline features, but that is a verification gap counted under demo quality, not a missing capability, so B and C's 9 is better supported.

No demo check was verified, the linked video is absent from the repo file tree, and a Remotion project in the repo can produce the footage programmatically, so real emulator recording cannot be confirmed. The described screenshots do appear to be real app captures, which supports 6 over A's more skeptical 5.

Pinned tool versions, env and test scripts, a prebuilt unsigned .hap with SHA-256 checksums, a per-feature verification table and the required AI disclosure docs. The gaps are macOS-first bash scripts, the Supabase dependency and no signed install on a real device, which justify the middle of A's 9 and C's 7.