jev-align

jev-align is an experimental CLI from Sutro for building AI Functions with decision models from TypeSafe, Perplexity, OpenAI, Liquid AI, and Cloudflare.

It finds uncertain examples, asks you to label them, and uses GEPA to improve the function. Use it in your application, keep learning from production examples, and share portable functions through ai-functions.dev.

Requires Python 3.11 or newer.

uv tool install jev-align
export TYPESAFE_API_KEY="..." # Or configure another model backend below
export OPENAI_API_KEY="..." # or ANTHROPIC_API_KEY / GEMINI_API_KEY
jeva

CLI aliases: decal, decision-align, jeva, jev-align. Install aliases: uv tool install decision-align or uv tool install decal.

Use pip install jev-align if you do not use uv. The guided setup discovers local CSV, Parquet, and JSONL files and includes three ready-to-run examples.

Each round:

1. Evaluates the configured dataset and measures uncertainty.
2. Selects ambiguous rows plus a random audit sample for you to label.
3. Uses your accumulated labels and optional rationales to run GEPA.
4. Shows the score, certainty change, and proposed definition diff.
5. Lets you accept, reject, rewind, or resume later.

The guided workflow asks you to review every label. Labels may be human-created, synthetic, imported, or agent-assisted; review them to the standard your use case requires. A higher training score never accepts a proposal automatically.

The guided Advanced menu configures:

• 5, 10, 15, or 20 training annotations per round.
• An optional 20% held-out evaluation set.
• GEPA's metric-call budget for guided labeling, which defaults to 300.

By default, jev-align uses the first 1,000 rows—or the entire dataset when it is smaller—and lets you concatenate all fields or select specific columns.

Everything can also be configured with flags:

jeva optimize posts.csv \
  --question "Is the post related to aviation?" \
  --column title \
  --column text \
  --pool-size 1000

Use repeated --class "NAME=DESCRIPTION" options for multiclass or multilabel tasks, and repeated --score-level options for scoring tasks. Run jeva optimize --help for the complete flag reference.

All backends support binary, multiclass, multilabel, and score tasks with text/JSON inputs. The guided setup detects configured providers.

Set the credentials above, then select a backend and optional model:

jeva optimize data.csv --question "Is this relevant?" --column text \
  --backend openrouter --backend-model cloudflare/clef

Cloudflare defaults to typesafe/jev-1.14.0 and also offers @cf/cloudflare/clef-flash. OpenRouter offers Jev, Perplexity Decider, GPT-6 Luna Decisions, Liquid D1.1, Clef, Clef Flash, and custom decision-model IDs using only OPENROUTER_API_KEY.

New runs default to TypeSafe Jev 1.14. Saved runs retain their backend and model; credentials come from the environment. --jev-model remains an alias for --backend-model. GEPA reflection is configured separately.

Backends preserve native probabilities and confidence and reject malformed responses. Transient failures receive up to four attempts, honoring server retry delays; delays over 60 seconds fail immediately.

GEPA's reflection model is separate from the decision model evaluating your data. The setup menu detects OpenAI, Anthropic, and Gemini keys.

New runs prefer OpenAI, then Anthropic, then Gemini. Saved runs retain their model. Use --reflection-model provider/model for any LiteLLM provider, including local vLLM:

export HOSTED_VLLM_API_BASE="http://localhost:8000/v1"
jeva optimize data.csv --question "Is this relevant?" --column text \
  --reflection-model "hosted_vllm/Qwen/Qwen3-8B"

• Arrow keys and Enter navigate menus.
• b returns to the previous label; /back leaves the rationale prompt.
• Space toggles choices in multilabel tasks.

Browse, run, or resume functions saved on your computer:

jeva functions
jeva optimize --resume .jev-align/runs/<run-id>

ai-functions.dev is the public registry for AI Functions built with jev-align. Browse shared functions on the website and publish your own from the CLI.

To publish, open your saved functions:

jeva functions

Select a function, then choose Push to ai-functions.dev. If needed, jev-align walks you through GitHub login. You can name the function, add a description, review what will be public, and confirm the push. It will appear at ai-functions.dev/<github-user>/<function-name>.

A published function includes everything another person needs to run it or continue improving it:

• The accepted definition and input signature.
• The runtime backend and learning configuration.
• Labeled inputs, labels, splits, and optional rationales.

It does not contain unlabeled source rows, local dataset paths, API keys, reflection-provider credentials, or an unaccepted proposal. The registry is currently public, so do not publish labeled data you cannot share. Human review is encouraged for important tasks, but the registry does not claim or verify that every published annotation was created or reviewed by a person.

Anyone can pull a public function without signing in:

jeva pull github-user/function-name
jeva pull github-user/function-name --version 2

Pulling needs no login. It verifies the immutable artifact and creates a normal saved AI Function under .jev-align/runs/, ready to run or continue improving.

Owners can remove a function from discovery and future public pulls:

jeva unpublish github-user/function-name

Unpublishing is reversible: versions and annotations are retained, and pushing the function again restores it. It cannot revoke copies that were already downloaded.

Load an AI Function in your application and capture useful production examples:

from jev_align import AIFunction

is_aviation = AIFunction.load(
    ".jev-align/runs/<run-id>",
    capture=True,
)

prediction = is_aviation(
    title="Airport expansion",
    text="A new runway opens next year.",
)

Later, resume the AI Function and label the captured examples. GEPA uses that feedback to propose the next version:

jeva functions

Squash is an intentionally dangerous, label-free demo that asks GEPA to reduce the evaluation model's uncertainty on a fixed dataset. Lower uncertainty is not correctness: a run can succeed by confidently assigning every row the same answer.

Select squash under Advanced, or pass --squash:

jeva optimize posts.csv --squash \
  --question "Is the post related to aviation?" \
  --column title \
  --column text

It supports every task type, preserves configured labels or score levels, uses an automatic metric budget, and runs up to five search passes by default. The final definition still requires approval. Squash records no annotations and cannot be published.

See AGENTS.md for detailed setup, provider configuration, CLI operation, and development guidance for coding agents.

Sutro is not affiliated with TypeSafe AI, the makers of Jev.