Built for Jev. Open to other models.

Run decisions.
Across your data.

Classify and score your whole dataset.
Connect stages. Trace every answer.

Interactive example · sample data

Find customers showing cancellation intent

Example complete
tickets.csvGrouped by customer

Swipe the map to explore →

6 customersINPUTSignals3 QUESTIONSFilterCODEDetailsNEXT STAGE

Every node is a record at a step. Select one to follow its path.

From a folder of files to a repeatable workflow.

CSVJSONJSONLTXTMarkdown

Ask together.
Go deeper after.

Give the model a record and a set of questions. Independent questions share one request.

Need an answer before asking the next question? That’s your next stage.

What makes a stage
STATE · CUSTOMER TICKETS

“Four outages this month. Fix this or we will not renew.”

Noul

Cancellation intent?

0.96probability of expressed intent
Choice

Main concern?

ReliabilitySupportPricing
Score

Business impact?

Critical work blocked
One request. Three typed answers.Illustrative output

Follow the whole story.

Which records made it through? What did the model see? Open a node and inspect the input, questions and raw response.

Keep the path.

Follow records through groups, filters and model stages. Excluded rows stay visible.

Inspect the evidence.

See the source data, typed results, and provider request and response behind a decision.

Change it. Run it again.

Refine questions or switch models. Start a new run without overwriting the earlier one.

YOUR DATAtickets.csv + accounts.json
Group by customer
STAGE 1Jev

Find cancellation signals

Intent · Concern · Impact

Keep matching customers
STAGE 2Kev 4B

Investigate the reasons

Selected customers + ticket history

Structured results

Start with a query.
Build from there.

Run a question set against your data. Save it as a stage, then connect it to the next one.

Choose a model at each stage.
Use hosted Jev or a supported self-hosted model, including Kev 4B.
Let code handle the data.
Group, join and filter between model calls.
Keep your inputs versioned.
Upload a fresh dataset. Earlier runs keep their original snapshot.
Build your workflow

A few practical details.

What data can I bring?

Upload CSV, JSON, JSONL, TXT or Markdown. That can be exported support tickets, articles, account records, or call and interview transcripts. Audio needs to be transcribed before upload. Current default limits are 20 MB and 10,000 records per file.

Do I have to use Jev?

No. Choose a supported model for each stage. Jevreduce supports hosted Jev and self-hosted models such as Kev 4B through provider adapters. The model you use determines which primitives and limits are available.

How is this different from asking a chatbot?

A run applies the same versioned questions and rules across a dataset. The platform stores structured outputs and the path each record took, so you can inspect individual decisions and rerun a workflow with new questions or models.

Can I use Pylon, Linear or Fireflies data?

Start with a supported file export from your tools. Native connectors are planned; this page’s examples do not require a live connection to those services.

Does replay give the same answers?

Replay creates a new run using the pinned inputs and workflow definition, with any supported overrides you choose. Model answers can change. The original run remains available for comparison.

Your data. Your questions.
Every path visible.

Start with one dataset. See where it takes you.

Open jevreduce