@elise_ai links an article by Mario Martone and Kyran Flynn at EliseAI. They tested @typesafeai Jev 1.14.0 experimental photo mode on whether an appliance serial-plate photo is readable enough to request a retake while the technician is still there. On 140 training photos, readability scores were comparable to larger models, and the checks were faster and a fraction of the estimated cost of Sol 6.1 and Opus 5.5.
- 10.7k views
- 20 likes
- 24 bookmarks
- 2 reposts
About this case
Product. Elise Labs (@elise_ai) posts only the article link, “How Typesafe AI’s Jev helps read appliance serial plates in the field.” Mario Martone, Head of Applied Research, and Kyran Flynn, Researcher, write that EliseAI builds AI for rental housing, reaching 1 in 5 U.S. apartments. A technician photographs a label inside an appliance instead of transcribing it. They labeled 140 training photos: 105 had at least one complete readable identifier and 35 did not. Each model scored whether a complete model number was readable and whether a complete serial number was readable. The higher score was the photo’s score, and a retake was flagged only if both were at or below 0.60. Jev answered both questions in one request. Sol and Opus used one call per question. ROC-AUC was 0.9989 for Sol, 0.9952 for Jev and 0.9951 for Opus. Jev’s median check was 94 ms, about 30 times faster than Sol’s 2.83 seconds, at about 2% of Sol’s estimated cost. At the 0.60 cutoff, Jev flagged 32 of the 35 unreadable photos and 2 of the 105 readable ones. Sol and Opus caught all 35 unreadable photos and flagged 10 readable ones each. The article says those 0.60 results are from the training set, not a held-out estimate, and that the costs are estimates from recorded token use and October 2026 rates, not invoices. Models named: Jev 1.14.0 (experimental photo mode), Sol 6.1, and Opus 5.5. No tweet media.
Original posthttps://x.com/elise_ai/status/2120965516072243219