udipta basumatari

This is the Jev explainer for business owners from a non-developer. I sorted 20,000 emails, slacks and transcripts last night into upsells, complaints, missed follow ups and 5 other things. It took 7 minutes and cost $1.45. The short version: Don't think of Jev as another model. Think of it as a tool for your models. Claude or GPT does the thinking and works out what questions are worth asking. Jev answers those questions across everything you've got. Fast, and for almost nothing. "Is the person in this email disappointed in our services?" It can answer a bunch of questions at once, too. ( something that current models can struggle while keeping accuracy.) "Are they disappointed" + "is there an upsell here" + "did we use the right opening line" + "is it a technical problem". Adding questions doesn't change the price much. Jev charges on input tokens - output is free. Long Version: So why does that matter for business owners? Because unlocking the next level of AI in a business isn't about a better model. It's about better data. The models are good enough now. The constraint is your data. Anything actually useful, spotting a problem before it lands, knowing which customers are about to leave, needs data that's been organised well enough to learn from. Most businesses don't have that. You've probably heard the terms structured and unstructured. Your emails, calls and messages are unstructured. A pile. Having the pile is one thing. Turning it into something you can count, trend and act on means putting a tag on every single item. That's the whole job Jev does, and it's the step most people skip because it used to be expensive. Two things I'm excited about doing with it. 1. Mass classification, for cheap. You can now tag every email, call and message you've got against as many questions as you want, and redo it whenever you change your mind. That's what makes a better second brain possible. Here's how. Everyone on Twitter is building a second brain. Plug in your emails, calls, docs, Notion, Obsidian, whatever, ask it anything, it knows your whole business. But it is always reactive. waiting for you to ask a question. To make it proactive - flagging stuff for you without you asking - something has to keep re-reading your data looking for things you care about, and that gets expensive fast. The cheap way, (and until now the only way), is to tag everything as it arrives so you can query the tags later. I tried that. Thought I was smart. You have to sit down and answer: what will I want to know about this data? What should I be tagging for right now? what connections do I have to make. And you'll get it wrong, because the technology is moving too fast to plan around and we are all building as we go. So you pick your ten tags, feel good about it. Six months later you launch a new service and want to know which existing customers have been asking for something like it. That tag doesn't exist, and nothing before today has it. Until now the choice was rerun the whole history through an LLM, which is a proper project with a proper bill, or live without it. Most people live without it. With Jev, retagging the whole history is a few dollars and a few minutes. So you stop trying to predict. Tag what matters today. When the business changes, retag everything. 2. Re-analysing old data for new patterns. If you've got a tonne of historical data, you can now structure it fast enough to actually learn from it. So you can use your tags (and retags) for pattern analysis. Say you've got three years of customer emails. Tag every thread for what the customer wanted, how we handled it, how it ended. Then ask what the ones that ended well have in common. Which way of communicating led to a good result, and which led to a bad one. I run a call centre, so here's mine. We do hundreds of thousands of calls a day and we do our best to classify them. Every call goes through an LLM and we ask it: "did they use the opener", "did they handle the objection", "did they ask for the appointment". Three or four things. Metrics I pulled out of my ass after a few years on the phones. It's what I think a good call looks like. Whether it actually is, I've never been able to check properly. At that volume, changing the questions and rerunning everything is a project I'm not equipped to run. Now I can hand the LLM the transcripts and the outcomes and say: come up with some hypotheses about what actually leads to appointments. It comes back with twenty. Did the caller use the prospect's name ten times. Did they mention a neighbour. Did they give a reason for calling in the first ten seconds. I point Jev at the history, reclassify every call against all twenty, line it up with which ones booked, and see if there's a pattern. Throw out the ones that didn't matter, ask for twenty more, run it again. None of that is new btw - It was just expensive and complex, and now it isn't. Both of these are the same thing underneath. You've got a pile of unstructured data, you've got questions about it, and the questions are going to change. Jev makes asking, and re-asking, cheap enough that you don't have to be right the first time which is great because businesses are more data driven than ever.