Jev in production › Message and ticket triage

openpoke-meets-jev (0xShin0221)

Fork of OpenPoke where Jev screens incoming email for importance and prompt injection, checks execution-agent tool calls, and filters search results before the LLM runs. Mean 424 ms per decision in an A/B against Claude Sonnet (author).

Open on GitHub ↗

424 msmeasured against a baseline, as published by the source
Use
Message and ticket triage
Industry
Productivity
Form
Open-source tool
Stage
Beta
Plugs into
OpenPoke
Listed
2026-09-20
Found via
github
Repository
0xShin0221/openpoke-meets-jev
Stars
1
Forks
0
Last push
2026-09-20
Language
Python
License
MIT
openpoke-meets-jev (0xShin0221) screenshot
docs/ab.png in the 0xShin0221/openpoke-meets-jev README, MIT; shown from GitHub.

The README opens with

A fork of OpenPoke that stops asking a chat model to make decisions.

claude-sonnet-4 jev-1.13.0 --- --- --- Latency, mean 2,452 ms 424 ms Latency, p50 / p95 2,512 / 2,883 ms 306 / 1,110 ms Input cost, 36 screens $0.1667 $0.0018 Input cost per 1,000 screens $4.63 $0.049 Distinct probabilities across 36 answers 6 18 Agreement at the 0.75 bar — 33 / 36

Read those as per decision, not per pipeline. The cost ratio is arithmetic on list rates ($3.00 against $0.042 per million input tokens) over unequal token counts — 55,554 against 41,739, because the LLM arm carries a tool schema. The latency numbers came off one machine in one afternoon, one provider per arm; I'd read 6× as an order of magnitude and not much more. There is no labelled set in this...

Badge

For the project's own README, linking back here:

Listed in Jev in production

Also used for message and ticket triage