Zulfi Tashleeh AI
Majmaah University · College of Science · Graduation Project

Al‑Zulfi SmartTashleh

Smart Search. Better Prices. Closer to You.

An AI‑powered scrapyard platform — Zulfi Tashleeh AI — that automates the appraisal and cataloging of vehicle wrecks in the Al‑Zulfi industrial zone. A dual deep‑learning pipeline detects and grades salvaged car parts from a single photo, then prices them dynamically against live market indicators, replacing the manual, unscientific workflow of traditional "Tashleeh" yards.

<1.5s
End‑to‑end inference latency target
0.962
R² score of the XGBoost pricing model
10 min
Avg. listing time, down from 3 hours
200
Concurrent uploads supported
A salvage yard packed with dismantled cars stacked side by side
AL‑ZULFI, KSA
About Us

Three students, one supervisor, and a yard full of unlisted parts.

We are a graduation‑project team from the College of Science at Majmaah University. Our project started in the Al‑Zulfi industrial zone, where a good spare part can sit unsold for months because nobody outside the yard knows it exists.

Zulfi Tashleeh AI is our answer to that. We combine computer vision and machine learning so a yard technician can photograph a wrecked vehicle and get its parts identified, graded, and priced in minutes, and so a driver or workshop can find the right part at a fair price without calling every yard in town.

We built the platform with the people who will use it. Yard operators in Al‑Zulfi tested it during a two‑week trial, and their feedback shaped the interface, the grading labels, and the way prices are approved.

Mission
Make every usable part in a Tashleeh yard easy to find, trust, and price.
Vision
A connected salvage market across the Kingdom, starting from Al‑Zulfi.
Values
Transparent prices, honest condition grades, and tools that work in real yard conditions.
Based in
Majmaah University, College of Science. Field work in the Al‑Zulfi industrial zone.
Project Goals

Five goals the platform is built to reach.

  1. 1

    Cut cataloging time from days to minutes

    Replace the manual inspection of more than 100 components per vehicle with a photo‑based workflow, so a wreck reaches the catalog the same day it reaches the yard.

    TargetAbout 10 minutes per listing
  2. 2

    Identify and grade parts from a single photo

    Detect each part with YOLOv8 and grade its condition as A, B, or C with ResNet‑50, giving buyers a standard quality label they can rely on before they pay.

    TargetResult in under 1.5 seconds
  3. 3

    Set fair prices from data

    Replace on‑the‑spot estimation with a pricing model that reads condition, model year, mileage, supply, and demand, so the same part gets the same price logic in every yard.

    TargetR² of 0.96 or higher
  4. 4

    Connect buyers and yards on one platform

    Let a driver or workshop search by name or photo, compare offers from several yards, and pick the nearest, cheapest, or best‑rated seller in one place.

    TargetSearch to offer in 5 steps
  5. 5

    Support the circular economy

    Raise the share of parts that are reused instead of scrapped, in line with Saudi Vision 2030, by making salvaged parts as easy to buy as new ones.

    TargetMore parts reused, less metal sent to landfill
The Solution

One photo in, a ranked list of offers out.

Imagine a user whose car has broken down and needs a spare part. They open the app, search by name or snap a photo, and the platform identifies the part, compares real offers, and connects them to the best seller — nearest, cheapest, or most trusted.

01

Search

Type a part name or capture a photo of the broken component.

02

AI Identifies Part

Computer vision detects and classifies the exact component.

03

Compare Offers

Prices and locations from multiple scrapyards are ranked.

04

Choose Best Option

The buyer weighs price, distance, and seller rating.

05

Order / Contact Seller

Place an order in‑app or message the seller directly.

01 — Vision

AI‑Powered Part Recognition

Identifies complex car parts instantly from a photo, replacing manual lookup with a single capture.

02 — Valuation

Smart Price Prediction

Predicts fair market prices from part condition, rarity, and real‑time demand — not guesswork.

03 — Discovery

Nearest & Best Seller

Recommends the closest scrapyard or the highest‑rated seller using geographic and reputation data.

User Experience

Designed for a dusty yard and a driver in a hurry.

Two people use the platform in very different situations. Each one gets a short path with a clear end.

Buyer: from a broken part to an order

5 steps

Open the app

No sign‑in needed to search.

Snap or type

Photo of the part, or its name.

See the match

Part name, grade, and fair price.

Compare yards

Sort by price, distance, or rating.

Order or call

Pay in the app or contact the seller.

Yard manager: from a wreck to a live listing

4 steps · about 10 minutes

Photograph the wreck

Works offline in the yard.

Review detected parts

Each part arrives with a grade.

Approve the price

One tap to accept, or edit it.

Publish

Listings sync when signal returns.

Readable in direct sunlight

High‑contrast colors and large type, because most listing happens outdoors at midday.

Big targets, one hand

Primary actions sit in thumb reach and are large enough to hit with dusty or gloved fingers.

Works without signal

Photos and drafts are saved on the phone and sync on their own once the connection is back.

Fast feedback

An appraisal returns in under 1.5 seconds, with a visible progress state while the models run.

The AI explains itself

Every grade shows the detected part and the reason for A, B, or C, so buyers can trust what they see.

People stay in control

No price goes live until the yard manager approves it, and any AI result can be corrected.

Arabic first

Full right‑to‑left Arabic interface with English available, and prices always shown in Saudi riyals.

Runs on modest phones

Built for devices with 2 GB of RAM and an 8 MP camera, the phones yard staff already carry.

Parts Marketplace

Find the part. Compare the offer.

Search the catalog, sort by price, distance, or condition grade, and add what you need to your cart. Pay with Apple Pay or a mada card.

Demo storefront: listings, sellers, and prices are sample data for the graduation project, and checkout does not process real payments.

Parts Price Guide

Approximate prices for common used parts.

Estimated prices in Saudi riyals for used parts in Al‑Zulfi yards, by condition grade. Sign in to save the parts you are watching.

PartCategoryFits Grade AExcellent Grade BGood Grade CFair

Prices are approximate estimates for guidance only. Actual prices change with the vehicle's model year, mileage, part availability, and the yard. Confirm the price with the seller before you buy.

Under the Hood

A dual deep‑learning pipeline, priced by gradient boosting.

Zulfi Tashleeh AI chains three models: one to find the part, one to grade its condition, and one to price it — combining computer vision with tabular market data.

STAGE 1 · DETECTION

YOLOv8

A real‑time object detection network localizes and segments core outer vehicle body parts directly from the uploaded image, drawing bounding boxes around each component.

Base architecture — YOLOv8‑Medium 150 training epochs · batch size 32 Optimizer — AdamW
STAGE 2 · GRADING

ResNet‑50

A secondary classifier network evaluates the physical condition of each detected part and grades it — Grade A, B, or C — for transparent, standardized quality reporting.

ResNet‑50, pretrained backbone 80 training epochs · batch size 64 Optimizer — SGD with momentum
STAGE 3 · PRICING

XGBoost Regressor

Classified metadata is combined with localized economic indicators — supply, demand, mileage, model year, and scrap‑metal index — to output a dynamic, defensible market price.

R² = 0.962 on held‑out test data MAE = SAR 180 · RMSE = SAR 290 Outperforms linear & random‑forest baselines
@app.post("/api/v1/appraise") async def appraise_part(file: UploadFile): # 1. Read uploaded image bytes img = read_image(await file.read()) # 2. Run object localization bboxes = yolo_model.predict(img) # 3. Classify condition grade grade = resnet_classifier.predict(bboxes) # 4. Calculate dynamic pricing suggested_price = xgboost_regressor.predict(grade) return { "detections": bboxes, "wear_grade": grade, "appraisal": suggested_price }
Client & Field Devices

React.js + Tailwind on a resilient edge

A high‑contrast, dusty‑daylight‑readable UI runs on standard Android/iOS phones (2GB RAM, 8MP camera minimum). A local SQLite cache lets technicians keep working offline in remote yards, syncing automatically once connectivity returns.

Backend & Cloud

FastAPI on GPU‑backed inference

An asynchronous FastAPI service exposes self‑documenting endpoints; heavy model inference runs on cloud GPU instances. JWT authentication and TLS 1.3 protect every session, targeting 99.9% uptime with active data replication.

System Design

Three actors, one relational core.

The platform's functional boundary is defined by three actors, each with a distinct set of responsibilities, sitting on top of a compact relational data model.

Y

Yard Managers / Dismantlers

Upload wreck images, approve AI pricing recommendations, manage inventory listings, and dispatch sold components.

B

Workshop Buyers / Mechanics

Search the catalog by vehicle type or part class, compare condition grades, pay securely, and verify compatibility.

A

Central Platform Admin

Monitors model performance, audits disputed grading, verifies legal ownership via government registries, and processes payouts.

DATA MODEL

Entity relationships, simplified

User
userID · PKnameemailpassword
Seller / Scrapyard
sellerID · PK, FKnamelocationphone
Buyer
buyerID · PK, FKorders placed
Order
orderID · PKdate / statustotalPricebuyerID · FK
Part
partID · PKcategorycondition (A/B/C)sellerID · FK
PartPrice
priceID · PKprice / datepartID · FK
Technology Stack

Built for reliability in the field.

Every layer was chosen for low deployment cost and dependable performance under dusty, bright, low‑connectivity yard conditions.

FastAPI (Python) React.js Tailwind CSS YOLOv8 ResNet‑50 XGBoost SQLite (offline sync) JWT + TLS 1.3 Cloud GPU inference Locust load testing
Experimental Evaluation

Tested against baselines, load, and real operators.

The pipeline was benchmarked on a 1,240‑image test split, stress‑tested up to 200 concurrent users, and evaluated by five Al‑Zulfi salvage‑yard operators over a two‑week UAT period.

Pricing Error by Model
Lower is better · Mean Absolute & Root‑Mean‑Squared Error, in SAR
Linear Reg. 1,240 1,890 Random Forest 420 680 XGBoost (ours) 180 290
MAE RMSE
Prediction Accuracy (R²)
Higher is better · Coefficient of determination, 0–1 scale
1.0 0.684 Linear Reg. 0.891 Random Forest 0.962 XGBoost
Part‑Detection Confusion Matrix
YOLOv8 on a 1,240‑image test split
PREDICTED Hood Door Hood Door 95.4% 8.8% 4.6% 91.2%
Diagonal = correct classification. Off‑diagonal confusion stays under 9% for both classes.
Latency Under Concurrent Load
Locust stress test · average request latency (ms)
1500ms target 310 480 890 1,420 10 50 100 200 users
UsersSuccess RateCPU Util.
10100.0%12.4%
50100.0%34.8%
10099.8%61.2%
20099.4%88.5%
Field Listing Time — Before vs. After
User Acceptance Testing · 5 Al‑Zulfi operators, 2‑week trial
Traditional manual listing≈ 180 min
With Zulfi Tashleeh AI≈ 10 min
Operators cataloged an incoming wreck roughly 18× faster using single‑tap AI appraisal versus manual entry.
Usability (SUS)
4.8 / 5.0
Single‑tap appraisal approval flow, rated by field operators under direct sunlight.
Impact & Honest Trade‑offs

A circular‑economy tool — with real limitations.

Digitizing the salvage sector carries clear upside for Vision 2030's circular‑economy goals, but the project's own evaluation surfaces real constraints worth stating plainly.

Advantages

Why it helps

  • Simplifies the search process and improves user experience
  • Saves time and effort for both buyers and yard operators
  • Provides accurate part identification through AI
  • Enhances price transparency and comparison across sellers
  • Diverts metal waste from landfill via more efficient recycling
  • Connects buyers and sellers in one unified platform
  • Extensible to future services such as maintenance
Limitations

What still needs work

  • High development cost from integrating multiple AI models
  • Requires large, localized datasets to train reliably
  • Image‑recognition accuracy is not always perfect
  • Depends on stable internet connectivity in the field
  • Needs buy‑in and cooperation from scrapyard businesses
  • Prices must be kept continuously updated to stay useful
  • Raises data‑privacy and security considerations
Environmental

Circular economy at scale

By increasing salvage efficiency, the platform directly supports Saudi Vision 2030's circular‑economy targets — helping divert hundreds of tons of metal waste from landfills annually.

Economic

Faster, fairer transactions

Automated pricing helps small‑scale yard owners increase sales velocity, reduce inventory loss, and offer transparent, defensible pricing to buyers.

FIELD CHALLENGES OBSERVED
Heavy dust & sand accumulation on parts Deformation boundaries confuse bounding boxes Older phone cameras struggle in low light
Future Research Horizons

What comes after Grade A, B, C.

NEXT

Augmented Reality disassembly guides

AR overlays that show dismantlers the optimal disassembly path for a given part in real time, cutting extraction time further.

MID‑TERM

Drone inventory scans

Autonomous drones with visual sensors to map and audit large salvage yards automatically, replacing manual walk‑throughs.

LONG‑TERM

Blockchain escrow & provenance

Smart‑contract escrow for secure transactions and immutable verification of a component's origin and ownership history.

The Team

Majmaah University · College of Science

HA

Haney Alqhyz

Project Supervisor
HW

Hamad Mohammed Al Washel

Student
YN

Yasser Mohamed Al‑Nassar

Student
YM

Yazid Sultan Al‑Mousa

Student

Sign in

Demo sign‑in for the project showcase. Accounts are kept in this browser only, so do not reuse a real password.