A small AI agent that connects people who lost something with the item someone found. Built for the MLH Hacktoberfest hackathon with open-source models on Hugging Face.
- Node X is a lost complaint (what the owner reports). Node Y is a found item (what the finder reports).
- The AI's job is to read both and decide if they are the same thing: yes / maybe / no, a confidence, and a one-line reason. No embeddings.
- The match graph draws a line from each complaint to each found item the AI links. Click a line to read why.
- Chat agent (Findy): talk normally; it saves reports, searches, and starts claims by calling tools.
- Ownership is plain code, not AI. The finder's hidden detail is checked by a fixed rule (3 attempts). The secret is never sent to the AI, and contact details are released only after a correct answer.
- If the AI or internet is down, the forms and keyword matching still work.
- Create a free token at https://huggingface.co/settings/tokens (allow Inference Providers).
- Double-click
run.bat. The first time it asks you to paste the token and saves it totoken.txt(git ignores this file). - The browser opens at http://127.0.0.1:7860. If the AI sleeps, double-click
check_ai.bat.
Not on Windows: pip install -r requirements.txt, put the token in token.txt, run python app.py.
- Ask: "What items have been found?"
- Say: "I lost my black Dell backpack in the library" and answer Findy's questions. Match #1 appears with the AI's reason and a line in the graph.
- Claim found item #1 with "panda keychain on the zipper" -> the reunited counter goes up.
Python standard library + huggingface_hub (InferenceClient), model Qwen/Qwen2.5-72B-Instruct (change with the MODEL_ID setting), SQLite, a tiny built-in web server.
See CONTRIBUTING.md. MIT licensed.