Gifty
An AI gift recommendation agent. It turns enriched, LinkedIn-style contact data into the top 3 personalised, real purchasable gifts per contact, each grounded in a live product, validated, and scored before a human signs off.
Run it locally
Needs uv for the Python backend and Bun for the frontend.
$ git clone https://github.com/grimmy-dev/gifty$ cd gifty$ ./dev.sh setup
$ # add API keys to backend/.env$ # (Anthropic / Gemini + Tavily)$ ./dev.sh
The frontend runs on localhost:5173 and the backend on :8000.
What it does
Personalised picks
Reads enriched, LinkedIn-style contact signals to surface the top 3 gifts per contact.
Grounded in real products
Every pick is backed by a live product found via Tavily web search, never invented.
Validated links
Each product link is checked to be live and relevant before it ever counts.
Confidence & risk
Each pick carries reasoning, a note, a confidence score, and a risk level.
Human in the loop
A review step gates unsuitable suggestions before they reach a user.
Two-stage pipeline
A fast model extracts signals, then a reasoning model searches and recommends.
Every result is inspectable. Open any contact to read its full LLM reasoning and process, then copy it or download the whole thing as a JSON file.
Want the full picture? The repo has a detailed write-up of the pipeline, a run trace, the architectural decisions, and the eval methodology in its DECISIONS.md and EVAL.md.