Vendors Selling Agent

A software specifically for vegetable vendors in the United States, which includes virtual image interaction and multiple functions. Using RL, GNN, DRL technology and is highly combined with MongoDB.

Comment

Inspiration

Large grocery markets rely on intuition every time their vegetables come on sale — how much to pick, how much discount? How much price will a truck cost? We wanted small stands to have the same AI tools big grocers take for granted. We wanted to give small stands with the same features, in our view.

What it does

The web dashboard VendorGroove Net is a software with Gemini-driven 2D avatar, which can interact and provide help through voice interaction. 1.Smart Restock. Q-learning agent recommending today's restock quantity, one LSTM anomaly detector predicting automatic price discounting,
2.Item Management. Graph Neural Network pointing to trending SKUs and competitive items 3.Group buying coordinator. It groups nearby vendors to shared truck manifests; one Midori, a Gemini 2.0 Flash AI assistant as a Live2D avatar.

How we built it

FastAPI + Uvicorn backend, MongoDB Atlas for geospatial vendor and order storage, three custom ML models written in pure Python (no PyTorch or TensorFlow), Gemini 2.0 Flash for chat, and a vanilla JS frontend using TailwindCSS, Chart.js, Leaflet.js, and PixiJS + Live2D.

Challenges we ran into

Providing ML model state in stateless HTTP requests, implementing Live2D without build pipeline, displaying raw model outputs (Q-values, anomaly scores) for non-technical users, and reporting group-buy results with scarce real-world data.

Accomplishments that we're proud of

Three state-of-the-art AI models without external ML dependences, a fully connected end-to-end stack with no placeholder screens, and real geospatial queries to the live MongoDB records.

What we learned

Tabular Q-Learning is a real-world training. Current engineering decides AI quality over model quality. Coordination problems (shared trucks) are more valuable than optimizations (better predictions).

What's next for Vendors Selling Agent

Persistent model state across sessions, JWT authentication, real LSTM training of USDA markets, PWA off-line, supply source API order, and cooperative network analysis of the entire vendor graph. Then there are many simple ideas and demonstrations that actually add a lot of backend code

Share this project:

Updates