ABOUT
DropPilot is a web application that answers one expensive question for a small e-commerce seller: what should I sell next? The user picks a few categories, a market and a budget; the app researches live marketplace data, scores each candidate against five weighted criteria, and returns a ranked shortlist of opportunity cards — each with a letter grade, plain metrics, the profit maths, the complaints buyers actually make, and a recommendation written for a business owner rather than an analyst. A full live run over three to eight products takes 20–90 seconds and costs a few cents.
THE PROBLEM
Product research is the highest-leverage decision in dropshipping and the one most often made badly. A typical seller copies a product from a course list or a social ad, spends $100–300 testing it, and finds out afterwards that the market was saturated, the margin too thin to absorb ad costs, or the product arrives broken often enough to eat the profit in returns.
The information needed to avoid that is public — it is simply scattered and slow to assemble. Whether interest is growing or fading needs interpretation, not just a chart. Judging how crowded a market is means manual counting with no clear threshold. Working out what you keep per sale means two sites and a spreadsheet. And nobody reads 400 buyer reviews per candidate to find out whether a product will be returned.
The goal was not to add AI to product research, but to compress hours of that work into one sub-minute request, in language a non-technical shop owner can act on immediately.
THE SOLUTION
The defining constraint was trust: a research tool that invents numbers is worse than no tool. So the architecture separates what a language model is good at from what it is not.
Numbers come from a deterministic rubric. Every figure a user sees is plain arithmetic over the gathered data — five weighted criteria covering demand, competition, profit, standout potential and reliability, producing a weighted total and a letter grade. It is fixed and identical for the same inputs every time, which is what makes the output defensible. The twelve-month trend is separately classified as rising, cooling, flat or spiked, catching a fad the seller would be arriving late to.
AI is given three narrow jobs. It clusters up to forty real buyer reviews per product into concrete complaint patterns, each with a rough share and a severity — high reserved for defects, breakage and safety, the failures a seller cannot fix with better copy — and those clusters feed two rubric criteria directly. It writes the client-facing verdict over the already-computed numbers, from a prompt that requires directness about weak margins, crowded markets and return risk. And a listing generator turns a chosen product into platform-specific titles, store copy, keywords and a suggested price.
Every AI call uses a structured schema and falls back cleanly on failure, so a failed call never produces an empty card. The model never sees a blank slate: it
explains computed numbers, so it cannot invent a margin or a score.
THE OUTCOMES
DropPilot replaces hours of scattered research with a single request, returning a shortlist a shop owner can act on: a grade, four plainly-worded metrics, what is in the product’s favour and what to watch, the top customer complaints, and a recommendation ranging from a strong pick to one to skip. It never says “launch this” — the final call stays with the seller, which is why cautions are written as bluntly as pros.
Cost controls — a six-hour cache, an enrichment cap and cheap-model routing for bulk reading — keep a full run at a few cents, against the $100–300 a single bad product test burns in ads. Two intended data sources were deliberately left switched off rather than half-wired, with the resulting approximations documented plainly rather than presented as measurements.
The project demonstrates the division of labour we apply wherever AI meets decisions with money attached: compute what can be computed, let the modelread what is too long to read and say what is too dry to say, and make every layer survive the failure of the one beneath it.
