Mule Mart

01 · the problem

Seniors and other students at Colby often throw away perfectly good things because they can't find anyone who wants them. The main place to post those things is Colby Now, the daily announcements email, but posts there are easy to miss and hard to find again. There's also no way to tell whether an item that was posted is still available. My team therefore wanted to build a better place for the Colby community to buy, sell, and donate things.

02 · what I built

• Led the four-person team as lead engineer: set the technical direction, decided what to build and in what order, and split the work across the team.
• Wrote most of the backend, a Flask app that stores data in PostgreSQL through SQLAlchemy.
• Built Mule Mart's semantic search feature, which finds listings by meaning instead of matching keywords. It uses SentenceTransformers (all-MiniLM-L6-v2) and PyTorch to store each listing's title and description as a vector embedding, and it ranks results by cosine similarity.
• Wrote the storage service for images, which live in a private AWS S3 bucket. Images are uploaded and loaded through presigned URLs, so large files never pass through the app server.
• Containerized the app with Docker.
• Wrote the three GitHub Actions workflows, which run the tests, linting, and the Docker build on every change.
• Set up automatic deploys through a Procfile, which also runs database migrations on each release.

03 · how it works

1. 01sign in

   Students sign in with Google or a password, using their Colby email.

   student@colby.edusigned in

   student@gmail.comPlease use your @colby.edu email address to sign in.
2. 02post a listing

   The seller's photos upload from the browser straight to S3 through presigned URLs, and the listing is saved to PostgreSQL.

   browserprivate S3 bucket

   photos, through a presigned URL

   browserFlaskPostgreSQL

   listing details

3. 03prepare for search

   When a listing is saved, its title and description are turned into a vector embedding with all-MiniLM-L6-v2 and stored with it.

   The North Face parka White down parka, size M. Kept me warm through two Maine winters.

   all-MiniLM-L6-v2

   [-0.001, 0.146, 0.078, 0.064, …, 0.011]

   384 numbers, stored with the listing

4. 04search

   When a student searches, their search term is turned into an embedding with the same model and compared with every listing's embedding. Listings are ranked by cosine similarity, and those scoring below 0.25 aren't shown.

   winter jacketsimilarity

   The North Face parka0.54

   Winter boots0.46

   Wool beanie0.43

   Patagonia fleece pullover0.43

   Knit gloves0.29

   below 0.25, not shown

   Futon0.21

   Mini fridge0.20
5. 05message and order

   Buyers message the seller in the app, and the chat checks for new messages every 2 seconds. Orders go to the seller to approve or reject. Once the seller approves an order, the listing is hidden from other buyers.

   Hi! Is the parka still available?

   Yes, it is. I can meet on campus tomorrow.

   order: pendingapproved

04 · key decisions

05 · where it stands

• ✓Live at mulemart.com for the Colby College community.
• ✓Students can search listings by meaning. For example, a search for “winter jacket” can find a listing titled “North Face parka”.
• ✓I'm rebuilding the backend as event-driven Go microservices for Mule Mart 2.0.

06 · technologies used

PythonLanguage

FlaskBackend

PostgreSQLDatabase

PyTorchAI

SentenceTransformersAI

AWS S3Storage

DockerInfrastructure