Sort thousands of rows with Jev

Jev is a small model built for one job: answering the same closed questions about many things. Is this company a fit? Which department is this person in? How senior are they, from 1 to 5? It gives a typed answer and how sure it is — nothing else, no text.

Your agent writes the questions. Jev answers them, row after row, for a fraction of what it costs to have a chat model read every row.

Who this is for. Anyone in a Tulina workspace whose agent has to sort, qualify or triage a list — a table of accounts, a CSV you just imported, profiles from a LinkedIn search.

Why use it

• Cheap at volume. A row usually costs a tiny fraction of a credit. A 10,000-row table is around a hundred credits (about €1), depending on how much of each row is sent.
• Fast. Rows are judged in parallel, on Tulina's side. A full table does not tie up your agent for hours.
• A probability, not just a label. Every answer comes with how likely it is. You decide where to draw the line, and leave the uncertain middle to a person or a stronger model.
• Nothing passes through the chat. On a table, rows are read, judged and written back on Tulina's side. Your agent only sees the totals and a sample.
• Safe to re-run. Rows that already have an answer are skipped, so a run that stopped halfway picks up where it left off and never pays twice.

What it is not

Jev does not write, summarise or explain. It cannot search the web or look anything up — it only sees what is in the row. If the row says nothing about company size, Jev cannot guess it: enrich first, then qualify.

Three kinds of question

• Yes or no — "Is this a B2B software company?" Answer: a probability, such as 0.92.
• One of several options — "Which department: sales, marketing, ops, finance, other?" Answer: the most likely option and its confidence.
• A scale — "How well does this account fit our offer, from 1 to 5?" Answer: a score such as 3.4, and its confidence.

You can ask several at once. Each one is answered on its own.

How to use it

1. Turn Jev on

In Tulina, open Connectors, find Jev and turn it on. There is no key to add: it runs on Tulina's.

2. Ask your agent

Talk to your agent as you would about any other task. Say which table, which columns matter, and what you want to know. For example:

Qualify the "Formnext exhibitors" table with Jev. Use the name, description and country columns. I want: is it a manufacturer (yes/no), which sector (automotive, aerospace, medical, other), and fit with our offer from 1 to 5. Our offer is 3D printing software for industrial teams.

Jev only knows what is in the row and in your questions. Put your context — who you sell to, what a good account looks like — in the request.

3. Check a sample first

Ask your agent to try the questions on a handful of rows and show you the answers before running the whole table. A trial writes nothing. Look at rows you already know the answer to. If they are wrong, reword the question and try again — this is where the quality is won.

4. Run the table

Once the sample looks right, the agent runs the rest. Each answer lands in its own column, with its confidence next to it, and a column recording which version of Jev answered. Missing columns are created for you.

What the table looks like afterwards:

A yes/no column holds the probability itself. A choice or a scale has its confidence in the _p column beside it. Brightform is the row to look at by hand.

Then sort or filter on the new columns like any other.

Lists that are not in a table

Jev also works on lists your agent is holding — profiles from a LinkedIn search, companies from a web search — without writing anything. The usual use is to decide before spending: keep only the profiles that fit, then pay to find emails or phone numbers for those alone.

Search LinkedIn for heads of operations at French logistics companies. Use Jev to keep only the people who actually run operations, then find emails for those.

What it costs

Jev is billed on what it really costs: 2 credits for every $0.01 the model costs to run. Most rows cost well under a hundredth of a credit, so you usually see it in tens of credits, not thousands.

Longer rows cost more. Send the columns the question needs, not the whole record. The breakdown is on Settings → Plans, under Jev. See Credits and billing.

Your data

Each row is sent to OpenRouter, which passes it to TypeSafe, the company behind Jev. Send only the columns the judgement needs — leave out emails, phone numbers and anything personal the question does not use.

Troubleshooting

Answers look random. The question is too vague, or the row does not hold the information. Describe each option clearly, and check that the columns sent actually contain what the question asks about.

Many rows score low. Rows with little in them score low for lack of evidence, not because they are bad fits. Enrich them first.

Some rows were left empty. Rows with nothing in the chosen columns are skipped. Fill them, then ask your agent to run again — only the empty ones are judged.

A row shows jev_error in the version column. That row could not be answered as it is, usually because it is too long. Send fewer or shorter columns.