Contact

Data Analyst

Hi, I'm Falak

I turn raw data into clear answers.

Computer Science graduate in Riyadh, working with Power BI, Python, SQL and R. I clean the data, build the model, and explain the result in plain English.

The dashboard is the last step. Most of my work is getting the data and the model right, so the numbers can be trusted.

96K+delivered orders analyzed in Power BI
97%best accuracy across three image models
1,000customer reviews classified with NLP in R
4.57/5GPA, BSc Computer Science, Princess Nourah University

Works.

How do Brazilian shoppers pay, and does paying in installments change anything?

#PowerBI#DAX#DataModeling
96K+delivered orders
51.5%paid in installments
~15Mtotal sales (BRL)
4.16 / 5average review score
Credit cards drove about 80% of revenue
Sales by payment type, delivered orders, in millions of BRL (rounded)
View as table
Payment typeSales (BRL)Share of ~15M
Credit card~12M~80%
Boleto~2.8M~19%
Vouchers and otherSmall remainder~1%
E-Commerce Installment Payment AnalysisData: 2016–2018

I analyzed the Olist Brazilian e-commerce dataset from Kaggle to understand how customers pay, and whether paying in installments relates to order value and satisfaction.

Just over half of delivered orders were paid in installments. Boleto, a Brazilian bank-slip payment, came second to credit cards at about 2.8M, and vouchers were small. Monthly sales grew steadily from late 2016 into 2018.

Role: Analyst

Power BI, Power Query, DAX and data modeling, from raw tables to the finished report.

I rebuilt the data model from scratch
Power BI had auto-detected wrong relationships, so I removed them and built a star-style schema with many-to-one, single-direction links. Simplified view.

Clean first.

I cleaned and typed the data in Power Query, and fixed the headers of the category translation table.

Date and plan.

I created a custom Date table and a calculated "Payment Plan" column that labels each order as Installments or Paid in full.

DAX measures.

Total Sales, Total Orders, % Installment Orders, Avg Order Value, Avg Installments and Avg Review Score.

One clean filter.

A report-level filter keeps delivered orders only, so every number on the page counts the same thing.

A year-end "Wrapped" for YouTube, built from a raw activity log.

#Python#Streamlit#Plotly
Viewing heatmap
Day of week against hour of day

Illustrative pattern showing the chart's layout. It is not real watch data, because the dashboard only runs on your own machine.

YouTube History WrappedRuns 100% locally

YouTube has no on-demand yearly recap, so I built a dashboard that reads a personal Google Takeout export and shows top channels, monthly activity and viewing heatmaps for any year.

The export is a log of clicks, with no record of how long anyone watched. So the dashboard states its assumptions and offers two ways to count.

Role: Builder

View on GitHub

Parse the log.

I extract each event from watch-history.html with BeautifulSoup and load it into pandas.

State the assumption.

Exact watch time is missing, so "Estimated Hours" assumes 10 minutes per video, and the project documents that openly.

Two ways to count.

A toggle switches between Total Views, which includes re-watches, and Unique Videos, which removes duplicates.

Private by design.

Everything runs on the user's machine, so personal history never leaves it.

Teaching a model to tell a good restaurant review from a bad one.

#R#NLP#SentimentAnalysis
From the dataset
Wow... Loved this place.Liked = 1
Crust is not good.Liked = 0
The pipeline
1,000 reviews Clean text TF-IDF Logistic regression Confusion matrix
Restaurant Review Sentiment AnalysisNov 2023 – Feb 2024

I built a binary sentiment classifier in R that labels restaurant reviews as positive or negative, then visualized which words customers use most.

The dataset is evenly split, with 500 positive and 500 negative reviews, so the model has no majority class to lean on.

Role: Analyst

View on GitHub

Clean the text.

Lowercased the reviews, removed punctuation and filtered stopwords with tm and dplyr.

Weight the words.

TF-IDF turns each review into features and gives more weight to words that carry meaning.

Train and test.

An 80/20 split keeps test reviews unseen, and caret's confusion matrix evaluates the predictions.

Show the language.

A ggplot2 bar plot of frequent terms and a wordcloud2 word cloud make the results easy to read.

Raskalah. Comparing three vision models to make recycling easier in Saudi Arabia.

#ComputerVision#ModelEvaluation#GraduationProject
DenseNet121 reached the highest accuracy
Accuracy by training epochs after fine-tuning, on 2,467 images in six classes
View as table
ModelBefore training20305060 epochs
DenseNet1219%94%97%97%84%
MobileNetV216%91%88%81%94%
ResNet5012%34%84%78%84%
Smart Waste Classification AppSep 2023 – May 2024

Raskalah is a mobile app that identifies waste from a photo, points people to nearby recycling centers, and rewards them with points.

We tested three pre-trained CNNs before and after fine-tuning. DenseNet121 held 97% at 30 and 50 epochs and fell to 84% by 60, so the number of epochs changed the result.

Role: Team of four, graduation project

Princess Nourah University. Built with Flutter, Firebase, TensorFlow and Keras.

View on GitHub

Prepare the images.

A Kaggle dataset of metal, glass, cardboard, paper, plastic and trash, with standardized sizes and augmentation.

Compare fairly.

DenseNet121, MobileNetV2 and ResNet50, each measured before and after fine-tuning on the same data.

Ship the winner.

The most accurate model went into the app, so users get a waste category from one photo.

Keep people using it.

Points, rewards and company-sponsored challenges, with full Arabic language support.

More builds

Predicting position from voltage

Cooperative training at KACST

A predictive model that estimates the position of a mover in a free-piston linear generator from voltage measurements. I cleaned and analyzed sensor time-series data, compared model accuracy, and presented the findings to my supervisors.

LocalChat: private PDF assistant

PythonLangChainFAISSStreamlit

Chat with your own PDFs entirely offline. Vector search finds the relevant passages, and a small local model answers on an ordinary laptop CPU.

See the code

Bank management system

C++File storage

A console app for client records and transactions, with data saved to files.

See the code

About me

I'm Falak, a Computer Science graduate based in Riyadh. My work so far covers predictive modeling, statistics and reporting, and I also hold a diploma in marketing.

Today I support business partners at a leading GCC fintech, where I investigate issues, find the underlying cause and explain technical and financial matters clearly. In 2026 I received a Going Above & Beyond award for my performance, and my team won a Best Teamwork award.

I want to keep working on questions like the ones on this page, where a careful model leads to a clear answer.

Journey

  1. Partner SupportLeading GCC fintech, Riyadh
  2. Cooperative Training, AI and DataKing Abdulaziz City for Science and Technology (KACST)
  3. BSc Computer Science, GPA 4.57 / 5Princess Nourah University, Riyadh
  4. Diploma of MarketingFuture of Knowledge Higher Institute

My stack

  • Business intelligencePower BI, Power Query, DAX, data modeling, Excel
  • AnalysisPython (pandas, NumPy, scikit-learn), R (dplyr, tm, caret), SQL, statistics
  • Visualization and appsPlotly, Matplotlib, ggplot2, Streamlit
  • Machine learningPredictive modeling, NLP, computer vision, TensorFlow, Keras
  • LanguagesArabic (native), English (professional working proficiency)

Certifications

  • Certified Computer Science SpecialistSaudi Council of Engineers, 2025
  • Data FundamentalsIBM, 2024
  • Machine Learning with PythonIBM, 2024
  • Data ScienceMinistry of Communications and Information Technology, 2023
  • Oracle Database FoundationsOracle, 2022

Have a question your data should answer?