Getting started

Install

Jevstiller needs Python 3.10 or newer. CPU works out of the box; a GPU only speeds up the encoder.

It is not on PyPI yet, so install it from GitHub:

pip install "jevstiller @ git+https://github.com/tomerglick57/Jevstiller"                 # core: numpy only
pip install "jevstiller[jev,onnx] @ git+https://github.com/tomerglick57/Jevstiller"       # Jev adapter + ONNX encoders (CPU)
pip install "jevstiller[jev,torch] @ git+https://github.com/tomerglick57/Jevstiller"      # PyTorch encoders (CUDA if available)

extra	adds
`jev`	`JevTeacher`, via `typesafe-sdk`
`onnx`	ONNX Runtime encoders on CPU
`gpu`	ONNX Runtime encoders on CUDA
`torch`	PyTorch encoders from Hugging Face checkpoints

Try it without a key

A synthetic teacher stands in for Jev, so you can watch the whole loop on CPU in under a minute:

git clone https://github.com/tomerglick57/Jevstiller
cd Jevstiller
pip install -e .
python examples/quickstart_synthetic.py

The student starts at 0% of traffic, gets trained and shadowed, gets promoted, and then takes over most requests while the audit channel keeps checking agreement.

Put it in front of Jev

Set TYPESAFE_API_KEY in your environment, then:

from jevstiller import Task, Jevstiller, load_encoder
from jevstiller.teachers.jev import JevTeacher

task = Task(
    name="support_router",
    instructions="Which team should handle this customer message?",
    classes={                              # descriptions are sent to Jev verbatim: they are the spec
        "billing":      "Charges, invoices, refunds, payment methods",
        "technical":    "Bugs, errors, integrations, things not working",
        "cancellation": "Wants to cancel, downgrade, or close the account",
        "sales":        "Pre-sales questions, plan comparison, quotes",
        "other":        "Anything that does not fit the categories above",
    },
    target_agreement=0.98,
)

js = Jevstiller(task, teacher=JevTeacher(), data_dir="./jevstiller-data", encoder=load_encoder("base"))

r = js.classify("Please cancel my subscription")
r.label        # "cancellation"
r.confidence   # 0.97
r.source       # "teacher" at first, later "student:v9"

Early on every request goes to Jev and becomes a training row. Once there is enough data, Jevstiller trains a student, runs it in shadow, and promotes it only if it meets the budget. You don’t call anything to make that happen.

Check on it

print(js.status().report())

Task: support_router   version 89a7438c2f7d   mode: cascade   audit rate 2%
Production: student:v9   Shadow: -
Requests: 11,083   student 69.8%   teacher 30.2%
Agreement with teacher (audit, n=415): 99.40% [98.46%, 99.84%]   target 98%   OK
  note: agreement with the teacher is not accuracy.

The agreement line shows the point estimate, its confidence interval, and one of OK, inconclusive, or BROKEN. See the guarantee for what those mean.

How it works: the loop, piece by piece.
Configuration: every knob and its default.
When to use it: and when not to.

Getting started

Install

Try it without a key

Put it in front of Jev

Check on it

Next