GreenlightX

Empty UI
Filling details
Diagnosis 1
Diagnosis 2

The Problem

Someone spent a year on a graphic novel. Hand-lettered pages, 3am illustration sessions, a story they genuinely believed in. They launched on Kickstarter with everything they had. It failed to fund on the last day. $400 short. That's not a rare story. 58% of Kickstarter campaigns don't make it. And most of them didn't fail because the idea was bad. They failed because nobody told them that their goal was 3x too high for their category. That 45-day campaigns statistically underperform 30-day ones. That their blurb was half the length of every comparable campaign that funded. The data existed. 300,000 campaigns worth of it. Funded ones, failed ones, the patterns hiding in plain sight.

Nobody was putting it in front of the creator before they launched. That's what GreenlightX does.

What It Does

Paste your draft campaign : the category, goal, duration, blurb. GreenlightX queries a live BigQuery database of 300,000+ real Kickstarter campaigns, finds the ones most comparable to yours, and tells you exactly where your plan diverges from what actually funds. Then it rewrites it.

Not a probability score. Not "good luck." A concrete rewrite, every change backed by a real number from real campaigns.

How It Works

You paste your draft
The agent queries BigQuery for funded vs failed benchmarks in your category
It identifies every gap between your draft and what actually works
It pulls real comparable campaigns that funded as evidence
It rewrites your goal, duration, and blurb based on the data
You get a side-by-side table : original vs rewrite, every change cited

The Fivetran Layer

Fivetran isn't a checkbox here. It's what makes the whole thing work.

A custom Connector SDK pipeline moves the Kickstarter dataset into BigQuery. The Fivetran MCP server lets the agent check and trigger syncs as part of its own execution. Every benchmark the agent cites comes from a live SQL query against that Fivetran-loaded table and not from the model's training data, not from a hardcoded prompt.

That distinction is the whole point. The agent isn't guessing. It's querying.

What I Actually Learned

The hardest part wasn't the code. It was making the agent genuinely agentic, getting it to call BigQuery as a tool before reasoning, not after. An LLM will confidently make up statistics if you let it. The entire architecture is designed around preventing that. Real data first, reasoning second.

The Fivetran → BigQuery pipeline was the spine. Everything else depended on it.

What Was Hard

Getting Vertex AI credentials to work on Cloud Run took longer than it should have. Writing queries that matched the actual dataset schema meant debugging in production. Scoping the whole thing to one day meant making hard calls about what to cut. The core loop — query, diagnose, rewrite — stayed. Everything else was negotiable.