🛒 CampusKart: AI-Powered Student Marketplace

A second-hand marketplace where college students buy and sell books, calculators, lab coats, cycles and hostel items. It has three AI features: personalised recommendations from search history, a fair-price predictor and a spam-listing detector.

Tech stack: Python · Flask · SQLite · scikit-learn · HTML/CSS/JS. It needs no internet connection or database server.

Live demo: https://campuskart-delta-lilac.vercel.app (demo data resets periodically, because Vercel gives each server instance a fresh copy of the database).

---

1. How to run

pip install -r requirements.txt
python seed.py        # creates campuskart.db with demo data (run again any time to reset)
python app.py         # open http://127.0.0.1:5000

Optional: python test_app.py runs automated tests of every page and flow. It resets the database, so run python seed.py afterwards.

2. Demo accounts

On the login page you can click any demo account to log in instantly.

User	Email	Password	What to show
Admin	[email protected]	admin123	Dashboard, AI-flagged listings, users, SQL explorer
Rahul Sharma (CSE, 3rd yr)	[email protected]	demo123	Searched dbms book, python, laptop… → gets CS books and keyboard
Priya Patel (Mech, 2nd yr)	[email protected]	demo123	Searched engineering drawing, drafter… → gets drawing tools and calculators
Karan Mehta (ECE, 3rd yr)	[email protected]	demo123	Searched arduino, multimeter… → gets electronics
Aisha Khan (IT, 1st yr)	[email protected]	demo123	Searched cycle, mattress, kettle… → gets hostel items and cycles
Neha Gupta (Civil, 1st yr)	[email protected]	demo123	New user with no history → sees "Trending" (cold-start problem)

Seller accounts (vikram, sneha, arjun, meera, rohan, ravi @college.edu) also use demo123.

3. Suggested presentation flow (about 10 minutes)

1. Login page: explain the demo accounts and that every student has a different history.
2. Rahul: the home page shows "Recommended for you", each card with a reason such as "Because you searched 'python'".
3. Log out, then Priya: same website, completely different recommendations.
4. Neha: no history, so the site falls back to trending items. Search "cycle" and go back home: the feed is now personalised.
5. Search a typo, e.g. arduno or calculatr: there's no exact match, but AI shows "Did you mean…".
6. Product page: AI price badge (Great deal / Fair / Overpriced) and Similar items.
7. Sell page: enter MRP, condition and age, then click ✨ Suggest price. Type "whatsapp me" in the title to see the live spam warning.
8. Buy an item: it disappears from listings (trigger). Log in as the seller to see the notification (trigger).
9. Admin: dashboard stats come from views, there are flagged listings to approve or reject, you can block a user, and the Database explorer shows all tables, views and triggers and lets you run live SELECT queries.

---

4. Database design

ER summary

USERS 1──N PRODUCTS N──1 CATEGORIES
USERS 1──N ORDERS N──1 PRODUCTS
ORDERS 1──1 REVIEWS
USERS M──N PRODUCTS  (via WISHLIST, composite PK)
USERS 1──N SEARCH_HISTORY
USERS 1──N PRODUCT_VIEWS N──1 PRODUCTS
USERS 1──N NOTIFICATIONS

Tables (9)

Table	Purpose	Key constraints
users	students and admins	PK, UNIQUE email, CHECK role/year
categories	9 item categories	UNIQUE name
products	listings	FK seller/category, CHECK price > 0, CHECK condition/status
orders	purchases	FK product/buyer, CHECK status
reviews	rating of a seller per order	UNIQUE order_id (1:1), CHECK rating 1-5
wishlist	saved items	composite PK (user_id, product_id)
search_history	every search (AI input)	FK user, ON DELETE CASCADE
product_views	every product opened (AI input)	FK user, product
notifications	alerts, mostly from triggers	FK user

Normalization (3NF)

• 1NF: every column is atomic (no lists stored in one column; wishlist is a separate table).
• 2NF: the only composite key is wishlist(user_id, product_id), and added_at depends on the whole key.
• 3NF: no transitive dependencies. For example, orders does not store seller_id, because it can be derived through products, and seller rating is computed by a view instead of being stored.

Triggers (6)

Trigger	When	What it does
trg_order_validate	BEFORE INSERT on orders	Aborts if the item is already sold or the buyer is the seller (prevents double selling)
trg_order_placed	AFTER INSERT on orders	Marks product sold and notifies the seller
trg_order_cancelled	AFTER UPDATE on orders	Puts product back on sale and notifies the seller
trg_wishlist_sold_alert	AFTER UPDATE on products	Notifies users who wishlisted a product when it is sold
trg_flagged_listing	AFTER INSERT on products	Notifies admins when AI flags a listing
trg_listing_moderated	AFTER UPDATE on products	Tells the seller their listing was approved or rejected

Views (4)

v_product_details (4-table join), v_seller_ratings (AVG + GROUP BY), v_category_sales (LEFT JOIN + SUM), v_admin_stats (scalar subqueries).

Other DBMS concepts used

• Transactions: every write uses with db:, which commits on success and rolls back on error. Buying is atomic: the order insert, the product status change and the notification either all happen or none do.
• Indexes: on products(status, category_id), search_history(user_id, searched_at) and others.
• Referential integrity: foreign keys with ON DELETE CASCADE.
• Security: parameterised queries (no SQL injection), hashed passwords (Werkzeug PBKDF2), and a read-only connection for the SQL explorer.

---

5. AI features (all in ai/)

AI 1: Recommendations (ai/recommender.py)

• Technique: content-based filtering with TF-IDF + cosine similarity.
• Each product becomes a TF-IDF vector of its title, category and description, using character n-grams so typos still match.
• User signals come from the last 10 searches, last 10 viewed items and wishlist. Newer signals get a higher weight (10% decay per step).
• A product's score is the best weight × similarity across all signals, and that signal becomes the explanation shown on the card.
• Cold start: a new user with no history sees the most-viewed items of the last 30 days.
• The same index powers "Similar items" and the typo-tolerant "Did you mean" search.

AI 2: Fair-price prediction (ai/price_predictor.py)

• Technique: Linear Regression on the listings in the database.
• Predicts the resale ratio (price ÷ MRP) from condition (one-hot), age and category (one-hot).
• Suggested price = ratio × MRP, rounded, with a ±10% range.
• The same model labels listings for buyers: Great deal / Fair price / Overpriced.

AI 3: Spam-listing detection (ai/spam_detector.py)

• Technique: Multinomial Naive Bayes on TF-IDF words and bigrams, trained on 38 labelled examples.
• Plus explainable rules: phone numbers or links, payment words (UPI, GPay, OTP), a price above MRP, and a price below 5% of MRP.
• A score ≥ 0.5 saves the listing as pending. A trigger alerts admins, who approve or reject it.

---

6. Likely viva questions

• Why SQLite? It's a full relational DBMS (ACID, triggers, views, FKs) with zero setup, which is ideal for a demo. The SQL is standard, so moving to MySQL needs only small changes (AUTOINCREMENT → AUTO_INCREMENT, datetime('now') → NOW(), and RAISE → SIGNAL in triggers).
• How do you prevent two people buying the same item? The BEFORE INSERT trigger checks the status inside the same transaction, so the second insert is aborted.
• Why TF-IDF and not deep learning? The data is small, TF-IDF is fast and explainable, and it needs no training time.
• What is the cold-start problem? A new user has no history to learn from, so we show trending items.
• Why predict a ratio instead of the price? Prices range from ₹100 to ₹50,000. A ratio is on the same scale for all items, so one simple model works for everything.
• How is search history privacy handled? Users can view and clear their history from the profile page.

7. Project report (IEEE format)

report/CampusKart_IEEE_Report.docx (editable) and report/CampusKart_IEEE_Report.pdf are in IEEE two-column A4 conference format. Before submitting, replace the yellow-highlighted placeholders (author names, department, college, guide) in the Word file and export it to PDF again.

All diagrams are in report/figures/ as 300-dpi PNGs: ER diagram, relational schema, architecture, DFD level 0 and level 1, use case, recommendation flowchart, purchase sequence diagram, and website screenshots. To regenerate them:

python report/make_diagrams.py        # diagrams
python report/take_screenshots.py     # screenshots (app must be running; run seed.py afterwards)

8. Project structure

campuskart/
├── app.py              Flask routes (auth, search, products, orders, admin)
├── schema.sql          tables, indexes, triggers, views
├── seed.py             demo data (users, 58 products, histories, orders)
├── test_app.py         automated tests
├── ai/
│   ├── recommender.py  AI 1: TF-IDF recommendations
│   ├── price_predictor.py  AI 2: linear regression pricing
│   └── spam_detector.py    AI 3: Naive Bayes spam filter
├── templates/          HTML pages (Jinja2)
└── static/style.css    styling (no external CSS framework)