Meet JEV: The Decision Engine Behind Our Search

Search Boxes Are Easy; Search Is Not

Most marketplaces put a box at the top of a page and call it search. You type a few words, the site matches those words, you scroll. That works right up until you write a sentence instead of a keyword. A real request sounds more like "a pilot for a 40-acre solar inspection near Pune, on a fixed budget" — and a word matcher reads that as a pile of unrelated tokens.

That gap is what we rebuilt. The parts of search where a judgement has to be made now run on a decision engine called JEV, and everything else stayed where it was. Here is what JEV is, what it does today, and what it means for you.

Where JEV Fits

Search is really two jobs. The first is understanding what you asked for: the service, the city, the budget, the certification, the years of experience. The second is deciding which of the results actually answer it — is this listing a better answer than that one, and is this drone model a weapon? Those answers are not sentences. They are a number, or one label from a short list. Deciding is what JEV does.

What JEV Actually Answers

JEV is deliberately narrow, and the narrowness is the point. A request carries one situation and a set of questions, and every answer comes back as one of exactly three things: a probability between 0 and 1 that a statement is true, a choice of one key from a fixed set, or a position on a defined scale. No free text, no prose, nothing written on the fly.

That is what makes it safe to use here. A model that can only answer "how relevant is this, from 0 to 4" cannot invent a plausible fact, cannot hallucinate a price, and cannot change its output shape between two identical requests. When we ask whether a drone name is a weapon, we get a probability we can threshold and a category we can show. When we ask whether a listing answers your query, we get a position on a scale that maps straight onto the relevance score you already see on a result.

Where It Runs Today

Two jobs in the product run on JEV. The first is the safety screen. Every drone model added to the catalogue is classified before it can appear on a pilot's profile or a service page: banned, or allowed, and why. Loitering munitions, missiles, armed strike aircraft and military-only airframes are refused; camera, agricultural, survey and industrial aircraft such as a DJI Mavic 3, an Agras T40 or a Foxtech AYK-350 are not. The classifier returns a probability plus a category, and the refusal a pilot reads is composed from that category, so the explanation can never drift from the decision behind it.

The second is ranking. When you search, JEV grades every candidate against your sentence — in parallel, not one at a time — and the order it returns is the order you see.

The Numbers

We re-measure before adopting a new release of the engine, because a probability scale can behave differently across versions. On the version we run today, the safety screen returns a verdict in about a third of a second, and a full page of results is graded in under a second. A search costs a fraction of a cent.

When It Fails, Nothing Breaks

An engine that sits in the middle of every search is a liability unless you decide in advance what happens when it stops answering. Ours fails open in both directions. If JEV is unreachable, ranking returns nothing and search keeps the order the results already had — you still get results, just not re-ordered. The safety screen reports "screening unavailable" and lets the name through rather than refusing a legitimate one.

There is no dependency chain to unravel either: switching a use case off falls back to the plain ranking rules that were always there, so the site keeps working whatever happens upstream, and ranking quality has one owner and cost has one line. An admin can switch the whole engine off, or a single use case, without a deploy, and every switch records who changed it.

Why a Decision Engine Beats a General One

A general model that writes prose looks like it could do this job, and it behaves badly at it: it is slower, it is billed by the word, and it can answer differently on Tuesday. Asking a narrow engine a narrow question is boring, which is exactly why it can be trusted: the answer is a number, the threshold is written down, and the test that proves it still holds is the same test we ran last week.

What This Means for You

If you are hiring, the practical difference is that the detail in your request is used: a budget, a minimum experience, a location, and the things you explicitly do not want are all part of the sentence the engine grades against. Every drone on the platform has passed the safety screen, and the order you see reflects relevance rather than the age of a listing.

If you are flying, the catalogue you are listed in is screened, and nothing switched behind the scenes can take the site down — at worst, results keep the order they arrived in.

The honest test is your own sentence. Post your requirement or browse drone pilots, and write the query the way you would say it out loud. If the right pilot comes back first, the engine did its job. And if you want to know what a project like yours should cost, start with what actually drives a drone-pilot quote.