ruvector learned engine
native · on the demo server · no API call
0.00s
- decided
- 0
- correct
- —
- median latency
- —
- network calls
- 0
Jev · typesafe.ai System One / ruvector · @ruvector/typesafe
Same tickets. Same questions. RuVector learned the job.
Jev's real API answers, replayed from its receipts, against a ruvector engine that learned from 136 labelled tickets and now answers without calling any API.
- Synthetic support tickets
- Frozen test split, n = 150
- Jev replayed from real receipts, 21 and 25 Sep 2026
- Jev latency includes the network round trip
01 The race
150 tickets. Three questions each. Go.
ruvector decides every ticket live, here. Jev's lane plays back each answer it really gave, at the latency it really took, four requests in flight, as in the capture.
Jev jev-latest
api.typesafe.ai · replayed receipt
0.00s
- decided
- 0
- correct
- —
- median latency
- —
- network calls
- 0
02 The learning phase
Watch the tickets find their teams.
Each dot is one test ticket, placed by the department probabilities the engine gives it. A sure answer sits on its corner. A confused one drifts to the middle.
03 Read the fine print
Where Jev is better than its score, or simply better.
Jev's urgency answer looks weak at a 0.5 cut-off, but it ranks urgent tickets well. Judge both engines on ranking too, and with a cut-off chosen on validation.
04 Teach it
Wrong? Correct it. It learns in milliseconds.
Type any ticket. ruvector answers all three questions live. If an answer is wrong, give the right one: the engine adds the example and answers again.
One engine serves every visitor, so lessons are shared. Anyone can press "Forget my lessons" to start it over.
Pick an example or type a ticket, then press Decide.
05 Receipts