I run Salesforge and host live sessions with people who build outbound for a living. My latest guest was Grisha, who heads GTM engineering at RevSculpt and is one of the Forge Experts I work with.
Grisha has spent hours testing Jev, the decision model from TypeSafe. He uses it to qualify leads, score signals and decode past campaigns before he scales them.
He cut about 18,000 UK Shopify stores down to 300 qualified accounts in roughly 10 minutes. He also classified 530 job posts for two cents.
This post covers each workflow he showed along with the questions the audience and I asked him. You can also watch the full recording here.
Jev is a decision model that TypeSafe released in September 2026. Grisha calls this type of model a decide model because making a decision is the only thing it does.
Claude or ChatGPT will write and explain things for you while Jev skips the writing and returns a typed answer with a probability attached.
Every request uses one of three question types:
Grisha used a simple restaurant review to show how the three types work together. The review said the food was great but the table waited 50 minutes and nobody brought drinks.
The confidence value is the part I care about most. When Jev is unsure about a row you can send it to a smarter LLM for a second look and keep everything else on the cheap model.
TypeSafe lists Jev at $0.042 per million input tokens with no charge for output, and the current version also reads PNG and JPEG images (TypeSafe docs).
Before getting into GTM, Grisha showed a few everyday uses that make the model easier to picture:
I get cold emails with CVs attached all the time. I would happily let Jev give each one a fit score before I open it.
Image input is new and Grisha has not tested it in a campaign yet. He did run a campaign for a website design agency where he rated every prospect site on visual design as well as technical structure.
Grisha listed four jobs Jev will not do:
So he splits the work in two. Claude gathers and prepares the data, and Jev makes the call on each row.
Jev was first to market, and the category has filled up fast since September. OpenAI shipped its own Decisions API and Parallel launched one with live web search built in.
Grisha counted close to 30 decision models on OpenRouter during the session. These are the three he discussed and how they compare:
Pricing from TypeSafe and the Parallel pricing comparison as of November 2026 (source).
Most of the decision traffic Grisha sees still goes to Jev, and it is his default for anything that does not need the web. When someone asks him which model to pick he gives the same answer he gives about data providers.
Test two or three on your own ICP and keep the one that holds up, because nobody can tell you in advance which one fits your use case.
His first workflow came from a client who sells to e-commerce brands. The client needed UK Shopify stores that ship small products and struggle with delivery, and some categories like glassware had to be left out.
He split the job into five passes so each one stayed cheap:
He also checked volume in that last pass. A store with 10,000 reviews and two delivery complaints tells you nothing about its delivery problems.
A store only qualified when delivery issues made up more than 10% of its bad reviews. That rule surfaced 100 delivery complaints across 43 stores.
The list went from about 18,000 stores to 300 in roughly 10 minutes of work.
His next example was bigger. Grisha pulled 25,000 waste collection companies across the US and Canada from Google Maps and had Jev qualify every one of them for $1.50.
That works out to six cents per 1,000 leads, and the whole run finished in under five minutes.
He estimated the same job on a GPT-6 class LLM would take close to an hour even with parallel workers. It would also cost three to four times as much.
That difference in speed and cost is why he reaches for Jev on any AI lead scoring task that comes down to labels.
Grisha also uses Jev to decide which B2B buying signals deserve attention. He splits signals into two kinds:
One signal on its own is rarely enough, so he stacks them. He pulls leads and signals from Leadsforge and has Jev judge whether each signal is relevant to the offer.
Jev scores irrelevant signals down, and an account with three relevant signals lands in tier A.
I asked whether he applies the same approach to custom signals. He builds those with Claude or Codex and has not added Jev to that workflow yet.
Social listening tools that track a keyword return plenty of posts that match the words without matching your ICP.
Grisha tracked posts about AI native product management and had Jev score each one on the content and on the author's title, company and location. Scoring 10,000 posts with Jev first costs far less than sending all of them to an agent.
Job titles alone can mislead you. Grisha's example was a client selling route planning software that saves fleets fuel.
A dispatch lead and a fleet manager both fit the ICP on title. Only one of the two job descriptions actually covered planning daily routes and schedules, which is the exact problem the software solves.
Jev read the descriptions and kept the relevant ones, qualifying 530 job posts from the last seven days for two cents.
He prefers a seven-day window because this signal works best while it is fresh. When volume is thin he widens it to two weeks or a month.
Raj asked about recency from the audience. In staffing a funding signal is gold on day one and dead by day 30, so does Jev account for that?
Jev only does that if you tell it to. Grisha writes the rule into the prompt so that a signal under seven days old scores tier A and anything older scores tier B.
I go deeper on stacking and timing in my guide to signal-based outbound.
Grisha also uses Jev after replies come in. Fast responses convert better and it gets hard to answer everyone fast when positive replies pile up.
I deal with that every day and it is a good problem to have. It is also why so many GTM engineers now build reply agents.
Grisha uses Jev as the first step in his reply agent. It labels each reply as interested, not now or wrong person before anything else happens.
Jev splits positive replies further because they ask for different things. Some want pricing and some want a meeting while others come with an objection attached.
The agent then drafts a response that matches the label. In Salesforge every reply already lands in Primebox, and an AI SDR can pick up the conversation from there.
The same idea works for inbound. One of Grisha's clients sent support tickets, existing customers and new prospects through a single form.
Jev now reads each submission and routes it to the right person. It also sorts spam and promotions out of inbound email.
Grisha uses Jev to look back at campaigns that already ran and work out why they performed the way they did.
Start by pulling every reply from your sequencer, including the negative ones. Grisha pulls them from Salesforge through the Forge MCP or the API.
Next he has Jev tag every email that went out. These are the elements he tags and the labels he uses for each:
Once every send carries tags you can group reply rates by each one and see which combination earned the replies. You cannot scale a campaign when you do not know which part of it worked.
The same method works for ads and content. You can tag ad creatives by hook, format and angle and compare them against results. LinkedIn posts and reels work the same way.
I asked Grisha how much data you need before the tags mean anything, because without enough volume you are tossing a coin. His rule of thumb is 1,000 sends per variant.
Teams with bigger budgets can test more, and in some markets even 5,000 sends can feel small.
Hitting that volume for every variant takes enough healthy mailboxes. I spread my own sends across Mailforge mailboxes that stay warm in Warmforge.
Grisha shared the patterns he found in his own campaigns:
You will not always have a lead magnet ready, so generic emails still have a place in the mix. Use them when you run out of something valuable to give.
He also had a warning on sequence length. Please do not send a thousand follow-ups, because they will not work.
The right number depends on your industry, offer and ICP. I broke down how many follow-ups to send in a separate post.
I asked whether he uses past data to pick a campaign structure for a new client. He does, and it works like building lookalikes.
If he already runs two manufacturing clients that generate a thousand replies a month, a third manufacturing client starts with whatever worked for them. His team keeps a dashboard of every campaign and reply per client, which makes that history easy to pull.
Startups are harder because they have no data yet. For them the first few campaigns produce that data.
Salesforge customers ask me which channel to start with all the time. Grisha's first step is checking how active the industry is on LinkedIn.
Some industries barely use it at all. In construction cold calling works far better than LinkedIn does.
He ran into this the same week while building a list of local HR companies in Alabama. One owner had a personal LinkedIn profile while his company had no page at all.
LinkedIn is worth testing with SaaS and GTM buyers. Even so his default is email first because it scales, with LinkedIn added as a step inside the sequence.
You can add more LinkedIn accounts and the volume still will not match email. I laid out the numbers behind cold email vs LinkedIn outreach if you want the full comparison.
I run both channels, and my team is testing WhatsApp outreach right now. WhatsApp needs its own messaging because it behaves nothing like email or LinkedIn.
Arseni asked where to find signals. Grisha pointed to three sources:
When I asked for his top signal, he picked a custom one. Custom signals take time to build with Claude and they usually convert best, especially in industries few people target.
Funding depends on what you sell. It is a strong signal for tender management software and a weak one for outbound services.
I pushed back on that. A company that raises money and then spends heavily on ads could be a good outbound buyer too.
Grisha agreed for some segments. In e-commerce he checks the Meta Ad Library to see how much a store spends on ads before reaching out.
His broader advice was to research why past customers bought, because signals only matter when they point to a pain point. Nobody buys without one.
He also likes to say that the list is the message. With a tight enough cold email list you sometimes do not need to mention the signal in the email at all.
Competitor followers are the signal I use most. Salesforge has more than 100 competitors, so I pull followers of those competitors and of the tools I integrate with in Leadsforge.
Those people already know the problem and the category. I open by mentioning that they follow a company and ask whether they use it.
When the answer is yes I pitch the switch. I only run this against competitors where my win rate is high.
Grisha happened to be running a competitor follower list the same week, with hiring and other signals stacked on top. He avoids naming the competitor in email and refers to the category instead.
On LinkedIn he asks an open question and stops there. Email is less conversational, so he adds the offer.
If you want more ideas to test, I keep a running list of the GTM signals that convert for me.
Jev handles high-volume decisions like lead qualification, signal scoring and reply classification. It returns a label or probability with a confidence score instead of generated text.
TypeSafe lists Jev at $0.042 per million input tokens with no charge for output. In Grisha's runs that meant $1.50 for 25,000 Google Maps leads and two cents for 530 job posts.
No. Jev has no web search and cannot find contacts. You source leads and signals elsewhere and use Jev to score them.
For yes or no checks and labeling at volume it is much faster and cheaper. Grisha sends rows where Jev has low confidence to an LLM such as gpt-6-luna for a second pass.
Grisha uses 1,000 sends per variant as a rule of thumb. Teams with bigger budgets can test more variants at higher volume.
Yes. The current Jev version accepts PNG and JPEG images in the request alongside text.





