~/work/mule-mart

Mule Mart

AI-powered campus marketplace

A full-stack e-commerce marketplace for the Colby College community where verified students buy, sell, and donate items, with AI-powered semantic search and real-time messaging.

Python

Flask

PostgreSQL

PyTorch

SentenceTransformers

AWS S3

Docker

Role

Full-Stack Engineer

Timeline

Oct 2025 → Present

Links

GitHub mulemart.com

Semantic

vector search over every listing

N+1 → 0

query bottlenecks eliminated

presigned-URL media uploads

mulemart.com

01 · the problem

Campus buy/sell happens in chaotic group chats and bulletin boards: no trust, no search, no structure. Mule Mart gives the Colby community a verified marketplace where keyword search isn't the ceiling. Searching “warm jacket for Maine winter” should find a parka, even if no listing uses those words.

02 · what I built

Designed a scalable REST API backend with Flask and PostgreSQL, with Google OAuth restricting access to verified community members.
Implemented AI-powered semantic search with SentenceTransformers (all-MiniLM-L6-v2) and PyTorch: listings are embedded as vectors and ranked by cosine similarity instead of keyword overlap.
Integrated AWS S3 presigned URLs so product images upload directly to object storage without passing through the app server.
Built an in-app real-time messaging system for buyer-seller communication, with notification triggers for unread messages.
Optimized SQLAlchemy queries to eliminate N+1 bottlenecks and improve response times, and containerized the app with Docker.

03 · the pipeline

1
authenticate
Google OAuth verifies campus identity before anyone can list or buy.
2
list
Flask REST API persists listings to PostgreSQL via SQLAlchemy; images go straight to S3 through presigned URLs.
3
embed
Each listing description is encoded into a vector embedding with all-MiniLM-L6-v2 at write time.
4
search
Queries are embedded and ranked by cosine similarity, giving meaning-based discovery layered on top of the relational store.
5
connect
Real-time in-app messaging links buyers and sellers, with unread-message notifications.

04 · key decisions

Hybrid relational + vector architecture

Rather than reaching for a dedicated vector database, I layered semantic search over PostgreSQL, keeping transactional integrity for orders and auth while still ranking results by meaning. Right-sized for a campus-scale platform.

Presigned URLs for media

Routing image uploads directly to S3 keeps large payloads off the API server entirely: cheaper, faster, and the backend stays focused on business logic.

Kill the N+1s before they kill you

Profiling listing pages exposed N+1 query patterns in SQLAlchemy relationships. Fixing eager-loading strategies measurably improved response times.

05 · where it stands

✓Live at mulemart.com for the Colby College community.
✓Meaning-based product discovery that keyword search can't match.
✓Open source on GitHub.