CHOOSEBusiness workflows
Open Source Jev Clone
Intikhab@AzamIntikhab
This project is an open-source clone of the Jev API, designed to outperform the original in its core function by offering offline execution, zero cost, and open weights, while achieving competitive accuracy and significantly lower latency.
Jev is a paid api. i open sourced a clone that beats it at the one thing jev was built for.
trained in 30 minutes on a free colab t4. scoreboard, on jev's own public benchmark:
calibration error: 0.144 vs 0.057
accuracy: 0.727 vs 0.697
latency per call: 239ms vs 60ms
cost per call: paid vs zero
runs offline: no vs yes
weights: closed vs open
five of six. the one i lost, i can explain.
typesafe built jev around calibrated confidence, a model that knows when it isn't sure. they invented a training method for it and named it in the launch post. that is the headline feature. that is the one a 150m model on a laptop beats by 2.5x.
if jev is on your timeline and you haven't looked: it only makes decisions. no prose, no generated tokens. you hand it a state plus typed questions and it answers all of them in one forward pass with a confidence on each.
the accuracy gap is 0.03 and it is not a mystery. 1,016 training examples. not model size, not epochs, i tested both and ruled them out.
i also published the baseline that beat my own model. frozen embeddings, logistic regression on top, zero fine tuning: 0.670. my fine tuned model: 0.624. it sits near the top of the readme, not buried.
what got past it was ensembling both, then one temperature scalar that took calibration from 0.156 to 0.057 without moving accuracy by a single decision.
all of it is yours. apache 2.0, open weights, one notebook you upload and run all. 150m parameters, runs offline, no api key, no network, no bill. train it on your own workflow and it gets better than anything general.
the failures ship with it. the schema key present in 10% of rows that killed training 300 rows in. the temperature fit that ran to T=400 and silently destroyed the confidence signal. the onnx export that built an invalid graph and raised nothing.