Public Projects
Passion projects let me showcase how I solve problems for different business domains.
Customer Segmentation & Recommendation Impact Analysis
RevoU — Dec 2025
Student Development Coordinator Full Stack Data Analytics
Retail & E-Commerce
A wine retailer relied on one-size-fits-all marketing despite having 2,240 customers with distinct demographics, spending patterns, purchase channels, and campaign responses.
Apply EDA, RFM segmentation, and K-Means clustering to uncover the most rewarding customer groups, then deliver prioritized marketing recommendations with a quantified impact analysis scored by business impact, feasibility, and ROI.
- EDA, RFM segmentation, and K-Means clustering on 2,240 wine retail customers to identify high-value segments.
- Prioritized top 3 marketing recommendations using a business impact × feasibility × ROI scoring framework.
- Built a Looker Studio dashboard for stakeholder self-service EDA.
- RFM segmentation identified Champions & Loyal Customers (34.7% of customers) as the highest-value group, spending 2× the overall average — with wine ($626) and meat ($364) dominating their basket.
- K-Means clustering revealed 4 behavioral clusters: Premium High-Value (20.5%, avg spend $1,476), Engaged Mid-Tier (21.3%, $927), Deal-Seeking (14.4%, $475), and Price-Sensitive Occasional Buyers (43.8%, $86).
- Priority impact analysis ranked Cross-Sell Wine to Meat Buyers (score 1,134) and Personalized Email Marketing (1,080) as the top quick-win strategies — both low-cost with high feasibility and 10–12% expected revenue lift.
Stack: Python, Pandas, Scikit-learn, Looker Studio
Skills: Exploratory Data Analysis, Machine Learning, RFM Scoring, K-Means Clustering, Impact Analysis, Marketing Strategy, Data Storytelling, Data Visualization
Depression Risk Assessment with Machine Learning
Dicoding Academy — Feb 2025
Project Manager, Data Scientist
Healthcare & Mental Health
Depression is rising globally across all age groups — WHO and INAMHS 2022 data show a surge in cases — yet early detection remains limited, especially in institutions that have not turned available data into practical screening support.
Identify the key factors driving depression risk, build a classification model to predict high-risk individuals, and deploy it as an accessible web app so institutions and individuals can act on early warning signals.
- Led a team of 4 to benchmark 5 ML models and identify the top 10 depression risk factors from survey data.
- Deployed a Streamlit web app for real-time individual depression risk assessment.
- Built an interactive dashboard surfacing training dataset patterns in demographics, lifestyle, and depression prevalence.
- Feature importance analysis ranked suicidal thoughts, academic pressure, and financial stress as the top 3 depression risk factors — ahead of sleep duration, dietary habits, and family mental health history.
- Five ML models were trained and benchmarked (Random Forest, Extra Trees, Logistic Regression, AdaBoost, SVC); Logistic Regression was selected as the best model with 85% accuracy using 10 features.
- The deployed Streamlit web app enables real-time individual risk assessment and a data dashboard, making the model accessible for institutions to support early mental health intervention.
Stack: Python, Scikit-learn, Streamlit
Skills: Exploratory Data Analysis, Machine Learning, Data Science, Web Development, Project Management, Data Storytelling, Data Visualization, Teamwork, Leadership
JKT48 YouTube Comment Sentiment Analysis
Dicoding Academy — Jan 2025
Data Scientist
Music & Entertainment
JKT48's 'New Era' special performance video — Benang Sari, Putik, dan Kupu-Kupu Malam — drew tens of thousands of YouTube comments reacting to its bold new concept, creating a large pool of raw, unstructured audience feedback.
Classify YouTube comment sentiment into positive, neutral, and negative to quantify audience reaction to the performance, compare lexicon-based and pre-trained BERT labeling, and train machine learning models that can predict the sentiment of new comments.
- Collected 22,996 comments via the YouTube Data API v3 and cleaned them down to 11,650 usable top-level comments through deduplication, text cleaning, stopword removal, and custom Indonesian slang normalization.
- Labeled sentiment with the ID-OpinionWords lexicon (supplemented by custom English word lists) and benchmarked it against pre-trained BERT labeling.
- Trained and compared four model + feature-extraction combinations — SVM, Random Forest, and a BiLSTM deep learning model — to predict comment sentiment.
- Overall reception skewed favorable: 50% neutral, 39% positive, and only 11% negative across 11,650 comments — with 'keren', 'keren banget', and 'new era' dominating positive mentions.
- Negative comments were the smallest group but the most engaging — highest average likes (20.71) and replies (1.92) per comment vs. 16.28 and 1.00 for positive — clustering around the provocative 'kupu-kupu malam' theme.
- A BiLSTM (5 epochs, batch size 32, max sequence length 20) was the best classifier at 93.44% test accuracy, beating SVM + TF-IDF (91.09%) and both Random Forest variants; lexicon labels produced more consistent results than BERT labels.
Stack: Python, Pandas, Scikit-learn, TensorFlow, YouTube Data API
Skills: Exploratory Data Analysis, Natural Language Processing, Sentiment Analysis, Machine Learning, Deep Learning, Data Science, Data Storytelling, Data Visualization
Customer Segmentation & Behaviour Classification — Istanbul Shopping Mall
Dicoding Academy — Jan 2025
Data Scientist
Retail
Retailers hold rich transaction data but few ready-made labels. Using the Istanbul Customer Shopping dataset — 99,457 transactions across 10 shopping malls (2021–2023) — the challenge was to discover natural customer groups from purchasing behaviour, then build a model that could automatically assign any new transaction to a group.
Run a two-stage machine learning pipeline: first use unsupervised clustering to segment transactions into behavioural groups, then train supervised classifiers to predict the segment label of unseen transactions — clearing a silhouette score of ≥ 0.55 and ≥ 87% classification accuracy.
- Profiled and preprocessed 99,457 clean transactions — frequency/ordinal encoding, MinMax scaling with down-weighted gender (×0.2) and payment (×0.3), and engineered features like recency_days, month, and value (quantity × price).
- Selected K-Means (chosen over DBSCAN and K-Prototypes) and ran a random feature search, landing on the price + category pair and k = 2 via the Elbow method to lift the silhouette score from 0.214 to 0.757.
- Labeled the segmented dataset and trained Logistic Regression and Decision Tree classifiers on a 70/30 stratified split to predict each transaction's cluster.
- Two behavioural segments: non-clothing buyers (65K transactions, 65%, avg. ₺576.8, top category Cosmetics) and clothing buyers (34K transactions, 35%, avg. ₺901.1, top category Clothing).
- Both classifiers hit a perfect 100% across accuracy, precision, recall, and F1 — but read honestly, this signals trivial labels: the clusters split almost entirely on the category feature, making the target effectively univariate.
- Key lesson: silhouette assumes Euclidean geometry that suits numeric data, not mixed numeric/categorical features — for mixed data, K-Prototypes/Gower distance and validation beyond a single internal score are the right call.
Stack: Python, Pandas, NumPy, Scikit-learn, Matplotlib, Seaborn
Skills: Exploratory Data Analysis, Feature Engineering, Clustering, Classification, Machine Learning, Data Science, Data Storytelling, Data Visualization
Olist E-Commerce Performance Analysis
Dicoding Academy — Dec 2024
Data Analyst
Retail & E-Commerce
Olist, Brazil's largest department store marketplace, generated a rich public dataset spanning sales, payments, logistics, customer reviews, and geolocation — but the raw, multi-table data held no straightforward answers to how the business was actually performing across these dimensions or where its operational weak spots were.
Answer 9 business questions across sales, payment type, logistics, customer satisfaction, and geolocation through end-to-end data wrangling and EDA, then communicate the findings via explanatory visualizations and an interactive dashboard.
- Wrangled 9 interrelated Olist datasets (~100K orders, 113K items, 1M geolocation records) — fixing data types, missing values, and duplicates.
- Ran EDA across sales trends, top/bottom product categories, payment methods, delivery accuracy, freight costs, review scores, and geographic distribution.
- Built an interactive Streamlit dashboard surfacing each business question through the most relevant chart type, including Folium geospatial maps.
- Sales grew strongly from Oct 2016 to Aug 2018; bed_bath_table was the best-selling category (10,953 orders) while health_beauty generated the most revenue (R$1.23M), and security_and_services ranked lowest across orders, revenue, and units.
- 90% of delivered orders arrived late versus their estimated delivery date (8% early, 2% on-time) — flagging a major logistics gap; credit card dominated payments at 75% (76,505 orders) with the highest average transaction value (R$163).
- Customer sentiment skewed positive — 58% scored 5 and 19% scored 4 (only ~15% negative) — while customers and sellers clustered heavily in southern and eastern Brazil.
Stack: Python, Pandas, NumPy, Matplotlib, Seaborn, Folium, Streamlit
Skills: Data Wrangling, Exploratory Data Analysis, Descriptive Statistics, Geospatial Analysis, KPI Reporting, Data Storytelling, Data Visualization
B2B Commerce Customer Retention & Cohort Analysis
RevoU x Telkom Indonesia — May 2024
Data Analyst Associate
B2B E-Commerce
Telkom Indonesia's B2B marketplace — built with the Ministry of SOEs, the Ministry of MSMEs, and LKPP to help MSMEs join government procurement — needed stronger customer loyalty in a competitive e-commerce environment.
Evaluate current retention performance through retention rate, customer lifetime value (CLV), and average revenue per user (ARPU), then surface trends and patterns to recommend concrete ways to improve retention — focusing on the highest-potential hyper-active and seasonal buyer segments.
- Ran monthly cohort analysis on 12,810 active buyers, focusing on the hyper-active (7,643) and seasonal (3,330) segments identified earlier via K-Means.
- Built retention-rate, CLV, and ARPU cohort heatmaps for 2023 to track how each monthly acquisition cohort behaved over time.
- Translated cohort patterns into segment-specific retention plays — loyalty programs, seasonal incentives, and proactive support.
- Hyper-active buyers' January & February cohorts were the largest and stickiest — the January cohort climbed to 87.7% retention by month 11 and dominated CLV (Rp 419.6B) and ARPU all year.
- Retention was weakest for the April–October hyper-active cohorts, yet overall retention rose over time across cohorts, peaking in November & December.
- Seasonal buyers made more first orders in H2 (July–September) with lower overall retention than hyper-active buyers; ARPU declined over time for both segments, with the steepest drop from January to February.
Stack: Python, Pandas, Matplotlib, Seaborn
Skills: Exploratory Data Analysis, Cohort Analysis, Data Storytelling, Data Visualization, Teamwork, Leadership, Project Management, Communication
B2B Commerce Customer Segmentation
RevoU x Telkom Indonesia — Apr 2024
Data Analyst Associate
B2B E-Commerce
The B2B Commerce team believed its one-size-fits-all treatment of buyers and sellers was leaving customer potential untapped and limiting segment-specific support.
Use EDA and K-Means clustering to segment active sellers and buyers, characterize each segment's behavior, and pinpoint the highest-potential groups for proactive support and retention.
- Segmented 6,661 active sellers and 12,810 active buyers from 2023 marketplace data using K-Means clustering.
- Engineered behavioral features — recency, order frequency, GMV, AOV, revenue, voucher usage, and buyer/seller reach — to profile each cluster.
- Defined 3 seller and 3 buyer segments and flagged hyper-active sellers & buyers as the top targets for retention.
- Sellers split into 3 clusters: Hyper-Active (4,324; avg 36 orders, GMV Rp 807M, recency 19 days), Big-Order Occasional (749; highest AOV at Rp 107.6M but just 1 order), and Seasonal (1,588).
- Buyers split into Hyper-Active (7,643; avg 20 orders, GMV Rp 480M), Seasonal (3,330), and Small-Order Occasional (1,837) — the only segment that meaningfully used vouchers (median Rp 25,000).
- Hyper-active sellers & buyers showed the highest revenue-growth potential, making them ideal targets for proactive support and retention; voucher usage overall was very low at 3.2% of orders.
Stack: Python, Pandas, Scikit-learn, Power BI
Skills: Customer Segmentation, K-Means Clustering, Exploratory Data Analysis, RFM Scoring, Data Storytelling, Data Visualization, Communication
B2B Commerce Sales Performance Dashboard
RevoU x Telkom Indonesia — Apr 2024
Data Analyst Associate
B2B E-Commerce
The B2B Commerce team needed a clearer way to monitor marketplace performance and give the sales team a comprehensive, story-driven view of the data.
Design a multi-page Power BI dashboard to (1) understand the general performance of the business and (2) analyze seller performance — surfacing prioritized KPIs through interactive, filterable visuals.
- Built a 3-page Power BI dashboard (Overview, Payments, Map) on 2023 B2B marketplace data for the sales team and its manager.
- Prioritized 6 core KPIs — GMV, Revenue, AOV, # Order, # Active Seller, # Active Buyer — with an interactive KPI selector and slicers for month, order status, seller category, and province.
- Modeled the data with custom measures and parameters, then documented dashboard usage and insight-generation demos.
- 2023 delivered Rp 10.94T GMV, Rp 73.7B revenue, and a Rp 26.78M AOV across 137K orders from 6,413 active sellers and 12.2K active buyers.
- Active buyers (12.2K) nearly doubled active sellers (6.4K), flagging an opportunity to recruit more sellers or grow buyers per seller; DKI Jakarta, Jawa Barat, and Jawa Timur led both sides.
- GMV, revenue, orders, and active users all trended upward through the year — peaking in December with dips around April–May and September–October; voucher usage was only 3.2% of orders.
Stack: Power BI, DAX
Skills: Exploratory Data Analysis, Data Visualization, KPI Reporting, Data Storytelling, Stakeholder Management, Communication
Daily Supply Chain & Logistics Performance Analysis
RevoU — Mar 2024
Data Analyst
Supply Chain & Logistics
Supply chain managers need to review past daily performance, but sifting through historical shipment records is time-consuming and rarely reveals the big picture — slowing down adjustments and proactive problem-solving.
Build a daily supply chain & logistics dashboard and demonstrate, on a single day's data, how it streamlines performance analysis and supports faster, data-driven operational decisions.
- Built a daily supply chain & logistics dashboard in Looker Studio by blending order, warehouse cost, and capacity tables from a ~9,215-order dataset.
- Prioritized 7 day-level KPIs — order volume, transit time, ship-ahead/late counts, weight, and cost — for at-a-glance monitoring.
- Generated insights and recommendations from a single day's data to demonstrate the dashboard's decision-support value.
- PORT04 acted as the network hub, handling 9,041 of 9,215 orders, ~29.3M of 29.5M shipped units, and 150,948 kg of weight — nearly all activity concentrated at one origin port.
- Operations skewed fast and early: transit processing time averaged 2 days (median 2, max 4) and shipments arrived ~2 days ahead of schedule on average, with very few significant late arrivals.
- The day totaled $15.6M in cost across 183K tons, but the dashboard explicitly flags that day-level snapshots need historical benchmarks and richer order detail before driving trend-level decisions.
Stack: Looker Studio, Google Sheets
Skills: Exploratory Data Analysis, Data Storytelling, Data Visualization
What Kind of Game to Develop to Get 1M USD Global Sales Within 1 Year?
RevoU — 2024
Project Manager, Data Analyst
Gaming
A newly established game studio — Sigma Game Studio — needed data-backed guidance on what type of game to develop for its debut release.
Identify the characteristics of games that achieved $1M+ global sales within 1 year of release (2016–2017 data), and investigate whether critic and user scores meaningfully impact commercial performance.
- Analyzed 2016–2017 global video game sales data to identify platform, genre, rating, and regional patterns of $1M+ successful titles.
- Built an interactive Tableau dashboard for stakeholder self-service EDA with cross-referenced filters.
- PS4 dominated successful recent releases with $62.21M total sales across 21 games, followed by Xbox One ($21.98M, 10 games) — only 4 of 9 active platforms produced any $1M+ title.
- Shooter was the top-performing genre ($40.55M total, avg $3.38M per game); Mature-rated games had the highest total sales while Everyone-rated games had the highest average sales per title ($3.54M).
- Critic and user scores showed very low correlation with global sales (r = 0.24 and 0.09 respectively) and were dropped from the regression model — quality scores alone don't drive commercial success.
Stack: Tableau, Google Sheets
Skills: Exploratory Data Analysis, Correlation, Issue Tree, Project Management, Data Storytelling, Data Visualization, Leadership, Communication, Teamwork
Berita2600 Social Media Content Word Cloud Analysis
BPS-Statistics Education & Training Center — Jan 2024
Data Analyst
Public Relations & Communications
Pusdiklat BPS published 183 social media content pieces ('Berita2600') throughout 2023, but the recurring themes and messaging patterns buried in all that text were never systematically examined — leaving the PR team without a clear read on what their content actually emphasized.
Understand the key themes of Pusdiklat BPS's 2023 social media content and identify concrete opportunities to improve content strategy for 2024, using word clouds and word/phrase frequency analysis.
- Scraped 183 social media posts ('Berita2600') from the Pusdiklat BPS blog using Simplescraper and cleaned the text in spreadsheets.
- Generated word clouds and frequency charts in Python after Indonesian stopword removal, surfacing the top 15 words and phrases.
- Translated the most frequent terms into thematic insights and content recommendations.
- After stopword removal the 183 posts held 15.3K words/phrases (3.6K unique, ~84 unique per post); the top 15 terms made up only ~10% of the total, indicating broad topical coverage.
- 'peserta' (241), 'pelatihan' (213), and 'kegiatan' (154) dominated — confirming consistent, participant- and training-centric messaging aligned with the org's mission.
- High-ranking 'Pembukaan Pelatihan' (67 mentions across 183 posts) flagged an over-reliance on opening-ceremony coverage, surfacing a clear opportunity to diversify content in 2024.
Stack: Python, Google Colab, Google Sheets, Simplescraper, WordCloud, NLTK
Skills: Exploratory Data Analysis, Natural Language Processing, Text Mining, Web Scraping, Data Storytelling, Data Visualization, Consulting, Communication
Samba Commerce Transaction Charts & Dashboard
RevoU — Sep 2023
Data Analyst
Retail & E-Commerce
Samba Commerce is a fast-growing Brazilian e-commerce launched in 2021, selling 15K products across 50+ categories nationwide, with transactions set as the company's main OKR for the year.
Develop company-wide, transaction-focused visualizations and dashboards in Tableau and Looker Studio so leaders and department managers can monitor transaction trends, key metrics, and OKR progress.
- Joined Orders, Order History, and Payments tables in Tableau and blended them in Looker Studio to model the full transaction lifecycle (~49.5K orders, Jan 2021–Jan 2022).
- Built 8 transaction visualizations — KPIs, monthly order vs payment-value trends, top-10 sellers, customer geography, and Pareto charts — plus a Tableau dashboard mockup.
- Rebuilt the dashboard interactively in Looker Studio with cross-filtering, state-to-city drill-down, and a date/state/city/category filter panel.
- From January 2021 to January 2022, Samba had 48K active customers, 1.7K active sellers, and 49.5K orders — averaging 3.8K orders/month at ~R$600K monthly value, trending up except a December 2021 dip.
- Orders were highly concentrated: the top 6 states (led by São Paulo, Rio de Janeiro, and Minas Gerais) drove 80% of all orders, while the top 10 sellers handled 16.7%.
- 90% of customers purchased just one item per order at an average order value of R$157.7, flagging clear cross-sell and basket-size opportunities — with Amapá, Alagoas, and Paraíba showing the highest average order values.
Stack: Tableau, Looker Studio, Google Sheets
Skills: Exploratory Data Analysis, Data Storytelling, Data Visualization, KPI Reporting
RevoBank Credit Card Customer Segmentation
RevoU — Aug 2023
Data Analyst
Banking & Financial Services
Over six months, RevoBank ran a fully bank-funded promotion for credit-card users, distributing unlimited RevoShop vouchers via email/SMS, with each redemption worth 15 reward points ≈ €1.5 and the bank earning a 2.4% margin per transaction.
Provide an overview of the promotion's performance, segment credit-card users by spending behavior and promo sensitivity, and discover opportunities to customize the rewards program for each segment while reducing promotion cost.
- Cleaned ~107,658 RevoShop credit-card customer records in Python — filtered merchants, removed missing values, fixed data types — and engineered promo-sensitivity, total-sales, age, and tenure features.
- Answered 12 business questions through EDA covering demographics, sales, promo cost vs revenue, and the effect of message frequency on spend.
- Ran K-Means clustering (MinMax scaling, k=3 chosen via elbow + silhouette), profiled each segment, and built tailored reward recommendations.
- RevoShop generated €56.6M in six months (62% promo-driven), with €456K in promo cost producing €845K in revenue (~85% ROI) and ~72K promo-sensitive customers.
- Average transactions rose with promotional messages up to ~6 messages, then declined with oversaturation — pinpointing an optimal messaging frequency.
- K-Means revealed 3 segments — Promo-Insensitive (33%), High-Value Buyer (45%, highest avg transaction €139.68), and Shopaholic (22%, highest frequency 7.58 and total sales €733.55) — so promos should focus on the High-Value Buyer and Shopaholic groups while the Promo-Insensitive group is better served with personalized banking instead.
Stack: Python, Pandas, NumPy, Scikit-learn, Matplotlib, Seaborn, Google Colab
Skills: Customer Segmentation, K-Means Clustering, Exploratory Data Analysis, Data Storytelling, Data Visualization
Cohort Analysis with SQL: TheLook eCommerce
RevoU — Jul 2023
Data Analyst
Retail & E-Commerce
TheLook, an eCommerce clothing site, faced a potential 2023 crisis and needed to optimize resources by reassessing low-growth product categories and customer retention.
Use BigQuery SQL to identify the product categories with the lowest growth over the past year (for potential resource reduction) and run a monthly retention cohort analysis to measure how many users return after their initial purchase.
- Wrote BigQuery SQL with window functions (LAG) and DATE_TRUNC to compute month-over-month inventory growth per product category across 2022.
- Built a monthly retention cohort query — first-purchase cohort, cohort size, and period via DATE_DIFF — and visualized retention as a heatmap in Google Sheets.
- Translated category-growth and cohort-retention findings into resource-optimization and customer-engagement recommendations.
- Clothing Sets had the most volatile and lowest overall growth (avg -3.82%, swinging from +54.17% to -62.5%), followed by Skirts, Suits, Pants & Capris, and Outerwear & Coats as the lowest-growth categories.
- Across all categories, February and September posted the lowest average monthly growth (-11.64% and -5.60%).
- Retention dropped most sharply between period 0 and period 1 (early churn), with first-month return rates of only ~3.5–8.4%; later 2022 cohorts (Oct–Dec) retained better, hinting at seasonal/holiday effects.
Stack: BigQuery, SQL, Google Sheets
Skills: Exploratory Data Analysis, Cohort Analysis, SQL, Data Storytelling, Data Visualization
Mr. Chu's Personal Insight: Correlation & Regression on Luxura
RevoU — Jun 2023
Data Analyst
Retail & E-Commerce
Mr. Chu — a loyal Luxura user, investor, and Adibi fan — asked the Luxura team for a personal insight into his spending so he could better manage his expenses on luxury products. The analysis reused the cleaned 500-customer dataset from the Marketing Prioritization for Luxura project.
Use correlation and multiple regression to find which variables most strongly relate to and impact Mr. Chu's Adibi spend, then predict his Adibi order value from a given set of characteristics.
- Built a correlation matrix across all numeric variables to find the strongest relationships to Adibi amount spent.
- Computed absolute standardized regression coefficients and ran stepwise multiple regression to isolate the single most impactful predictor.
- Validated the final model against the four regression assumptions and produced a point prediction with recommendations.
- Amount Spent on Celinna and Amount Spent on Balena correlated most strongly with Adibi spend (r = 0.9998 and 0.9825).
- Standardized coefficients and multiple regression identified Amount Spent on Celinna as the only statistically significant predictor — final model: Adibi = -0.2993 + 0.695 × Celinna.
- The model predicted an Adibi order value of ~46,263 USD for the given profile and satisfied non-multicollinearity, normality, and non-autocorrelation, but failed homoscedasticity — flagging room for refinement.
Stack: Google Sheets, XLMiner Analysis ToolPak
Skills: Exploratory Data Analysis, Correlation, Regression, Statistics, Data Storytelling, Data Visualization
Marketing Prioritization for Luxura eCommerce Top Brands
RevoU — Jun 2023
Data Analyst
Retail & E-Commerce
Luxura is a luxury-fashion e-commerce that exclusively sells three globally famous, closely competing brands — Adibi, Balena, and Celinna. With only one to prioritize for marketing, Adibi looked obvious after paying the highest partnership fee ($80M in Feb 2023) — but purchase performance had to be checked first.
Identify the brand that performed best in sales and customer engagement over the last three months so Luxura can strategically allocate marketing to maximize sales and market share.
- Cleaned a 500-customer, 20-variable loyalty-card dataset in Google Sheets — handling missing values, duplicates, data types, and Tukey's-fences outliers.
- Ran descriptive statistics and EDA (with the XLMiner Analysis ToolPak) and pivot tables across demographics, income, marital status, and purchasing power.
- Ran three two-sample t-tests (A vs B, A vs C, B vs C) on amount spent to compare brand purchase value.
- Order quantity was a near three-way tie (Adibi 33.5%, Balena 33.3%, Celinna 33.2%), but monetary value diverged sharply: Celinna led total spend (R$117.4K) over Balena (R$104.0K) and Adibi (R$81.4K).
- All three t-tests rejected the null (p < 0.05): Celinna's average spend was significantly higher than both Balena and Adibi, and Balena's was higher than Adibi's.
- Single and high-income customers drove the most orders and spend across all brands — recommendation: prioritize Celinna, improve Balena, and re-evaluate Adibi (highest fee but lowest monetary value).
Stack: Google Sheets, XLMiner Analysis ToolPak
Skills: Descriptive Statistics, Hypothesis Testing, Exploratory Data Analysis, Data Storytelling, Data Visualization
Root Cause Analysis on XYZ Company's High Attrition Rate
RevoU — May 2023
Data Analyst
Human Resources
XYZ Company, a 4,000-employee organization, faced a 15% annual attrition rate that left many positions to backfill from the market.
Frame the business problem and design a root-cause-analysis approach — DARCI roles, a SMART objective, an issue tree of potential root causes, prioritized hypotheses, and tracking metrics — to reduce attrition from 15% to 5-10% within one year.
- Defined a DARCI matrix and a SMART problem statement/objective to cut attrition from 15% to 5-10% within one year.
- Built an issue tree of 10 potential root causes spanning personal, workplace, organizational, and external/market factors.
- Turned the root causes into 10 testable hypotheses, prioritized 5 as high-priority, and recommended 5 metrics to track.
- 78% of attrition came from employees under 40, and male attrition was slightly higher than female — making age- and gender-segmented attrition/retention the top hypotheses to test.
- Job satisfaction, relationship satisfaction, and number of trainings per employee were prioritized as high-impact, actionable drivers, while salary and standard working hours were deprioritized as policy-constrained.
- Recommended 5 metrics — attrition rate, retention rate, average job satisfaction, average relationship satisfaction, and average trainings per employee — to monitor progress.
Stack: Google Slides
Skills: Exploratory Data Analysis, Root Cause Analysis, Issue Tree, Hypothesis Prioritization, Metrics Recommendation, Data Storytelling