A hackathon backend that turns a plain-language brief into a specific campaign. Almost all of the quality came from what got retrieved before the model was called.
Role
Built the backend.
Period
2024
Team
A hackathon team, one day.
I owned
The backend and the generation approach
Multi-source context assembly
The team owned
The interface
The pitch
7 hrs
To a working backend
1st
Place
Context assembly is the product, not the prompt.
The path
01Plain-language brief
02multi-source context assembly
03generated campaign
The problem
We had one day at a company hackathon to show that an AI product concept could actually work.
What I built
The generation was the easy half. The work was in assembling enough context to make the output specific: scraped public material about the brand, CRM data on the account, and competitive enrichment, composed into something the model could reason over.
What changed
It won, and the concept went onto the product roadmap. That is what a hackathon build can honestly claim, and it is also the ceiling on what it proves.
The more useful result was for me. This is where I learned where LLM output quality actually comes from, and it has changed how I debug every AI system I have built since.
The pattern
A model given a thin prompt writes a generic campaign, and a generic campaign is worth nothing to a company that already knows its own category.
That has held up since. When the output disappoints, the retrieval is usually where the problem is, and the prompt is usually where people look.
The call
Decided
Spend the day assembling context from brand material, account data and competitive enrichment, instead of on the prompt or the model.
Instead of
A tighter prompt and a polished demo, which is where most of a seven-hour build usually goes.
What it cost
No interface. The backend needed someone else's front end to be shown at all, and it would not have placed on its own.