I Built an Agent That Finds Free Food Near You

#webdev #programming #qwen #ai

Most event discovery tools answer the wrong question.

If you search for events in a city, you usually get a long list of meetups, talks, mixers, launch parties, and workshops. That is useful, but it does not answer the question a lot of people quietly care about:

Which of these events are actually likely to have free food or drinks?

That is the idea behind Agent Free Food, my Build Small hackathon project.

It is a Gradio app that lets you ask for free food or drink events in any city. The agent searches the web, checks event pages, scores each result, and returns a ranked table of the strongest opportunities.

Try it here: https://huggingface.co/spaces/build-small-hackathon/Agentfreefood

Demo video: https://youtu.be/MMg47HF4oVA

Why I Built It

Free food is usually hidden inside event pages.

Sometimes it is explicit: “pizza provided,” “free drinks,” or “refreshments included.” Other times it is implied by the type of event: startup launch parties, hackathons, demo nights, mixers, happy hours, and community events often include food even when the title does not say so.

Searching manually takes time. You have to open many pages, read the details, and decide what is worth checking.

I wanted an agent to do that triage automatically.

How It Works

Agent Free Food uses Exa for live web search and a small hosted Hugging Face model for scoring.

The model does not browse the web by itself or invent events. Instead, the app gives it a narrow job:

Read event page text
Extract useful details
Estimate free-food likelihood
Return structured results

Each result includes the event name, score, location, food or drink evidence, and a short reason.

What I Learned

The biggest lesson was that small models work best when the system around them is clear.

The model does not need to solve everything. Exa handles search. The app handles routing, filtering, and display. The model handles judgment over messy event text.

That made the project more reliable and easier to evaluate.

I also used Codex throughout the build, including prompt tuning the agent instructions so it focused on strong free-food opportunities instead of generic event listings.