jevify/ 0.3

Jev-like answers,
but on Huawei Cloud MaaS

Jevify is a proxy that accepts Jev requests, runs them through DeepSeek V4.1 Flash on Huawei Cloud MaaS, and returns Jev-shaped answers.

Deployment instructions Try the playground
Request context

“I was charged twice.
Can I get a refund?”

j·
jevifyHuawei MaaS · DeepSeek
ChoiceAnswer
billing 0.95
technical 0.04other 0.01
Illustrative output

How Jevify works

Jev is TypeSafe’s decision model. Instead of free-form text, it returns a Choice from your options, a Score on your rubric, or a Noul: the probability that a statement is true.

Jevify implements that request and response format using Huawei Cloud MaaS. Change your API URL and key; keep your request body and answer handling. The proxy validates DeepSeek’s output and calculates scores and confidence in code.

The same request to both APIs

Save this as request.json. Both calls below send this exact body.

request.json

{
  "model": "jev-latest",
  "state": "Please refund the duplicate charge.",
  "questions": {
    "team": {
      "type": "choice",
      "instructions": "Which team should handle this?",
      "criteria": {
        "billing": "Payments and refunds",
        "technical": "Software bugs"
      }
    }
  }
}

Jev TypeSafe key

curl https://api.typesafe.ai/v1/systemone \
  -H "Authorization: Bearer $JEV_API_KEY" \
  -H 'Content-Type: application/json' \
  --data-binary @request.json

Jevify Huawei MaaS key

curl https://jevify.hwctools.site/v1/systemone \
  -H "Authorization: Bearer $MAAS_API_KEY" \
  -H 'Content-Type: application/json' \
  --data-binary @request.json

The same answer fields

answers.team.choice
answers.team.probabilities
answers.team.confidence

Values can differ. Jevify identifies its actual model as deepseek-v4.1-flash; its probabilities are LLM estimates, not Jev’s calibrated probabilities.

Playground

Edit an example and run it. New prompts use the server’s Huawei MaaS key; cached runs replay their original response time.

Request

ResponseReady

Response

Run an example to see what comes back.

Edited prompts are sent to Huawei MaaS. Please use sample data. 6 live runs / 15 min · cached for 24h.

LLM-estimated probabilities. Jev’s calibration is not reproduced.

↗

Choice enum + probabilities

One of your options. Up to 255, with a probability for every one.

≋

Score number + rubric

A weighted position on your ordered levels. Fractional, from 0 upward.

◒

Noul P(yes)

A number between 0 and 1. You set the threshold for yes.

Deploy Jevify

Download source ↓

You need Docker with Compose and a Huawei Cloud MaaS key with DeepSeek V4.1 Flash access. The API uses the caller’s key, so no server key is required.

1. Build and start the container

Terminal
The proxy API listens at http://localhost:3000. The health endpoint should return {"status":"ok"}.

    
2. Send a request with your MaaS key
cURL
Set MAAS_API_KEY in your shell. Jevify forwards that key only to the configured Huawei MaaS endpoint.

    
3. Point your Jev client at the proxy
TypeScript · @typesafe-ai/sdk
Install the SDK with npm install @typesafe-ai/sdk. Change the base URL and use your MaaS key. Existing question and answer code stays the same.

  




Jevify vs Jev: code and reasoning
A focused suite of 24 code and multi-step reasoning problems, selected to explore a general LLM’s strengths. Both models receive the same questions.
Loading measured results…