Rafael Widjajahakim

Hello World of Website Building

Thu, 01 Jan 2026 00:00:00 +0000

BACKGROUND STORY

Creating a portfolio website has been a personal project todolist for so many years. But, there were a lot of obstacles (mostly personal motivations) to clearly make one.

The more legit reasons are:

As many typical website builders (Wix, Squarespace, etc.) will require cost to host, which one is worth it to try for my very early stage build? Most of these are for eCommerce, and I don’t need many of these tools.
Is there a pre-made theme among these websites that fit to my need?
How to organize the website for what I can present/provide? I am not in Art & Design field where I have many pictures or showcase products.

A MORE ADVANCED PATH (why run before walk)

Sometime ago, I found portfolio website from someone in tech. The theme was not what I have seen in common website builders. At the footer of the page, it says ‘Built with Hugo, through Netlify". At that time, I have not coded with Python, Linux, or anything. I was not even familiar with Github etc. I just want to use that theme.

So, I tried to follow any articles I could find. But I could not finish testing and got a job that was not that technical anyway. So I thought, I wouldn’t need it anyway; so I stopped.

NOW

Fast forward, I found out that Github can host a simple static website through the pages. So, I guess it is time to finally check that todolist… except,…

MORE HEADACHE

{::comment}comment {:/comment}

ML Pipeline: Predicting with Scikit-Learn

Sun, 01 Mar 2026 00:00:00 +0000

Overview

This post documents an end-to-end supervised machine learning workflow built in Python during my graduate coursework at Boston University. The goal was to build a reproducible pipeline that handles data preprocessing, feature selection, model training, and evaluation — the kind of workflow that translates directly to real-world analytics problems.

Problem & Dataset

The analysis used a structured tabular dataset with a mix of numeric and categorical features. The target variable was a binary classification outcome. The challenge was to:

Handle missing values and skewed distributions thoughtfully
Engineer features that improve signal without data leakage
Select a modeling strategy that generalizes well

Pipeline Design

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28


from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler, OneHotEncoder
from sklearn.compose import ColumnTransformer
from sklearn.ensemble import RandomForestClassifier, GradientBoostingClassifier
from sklearn.model_selection import StratifiedKFold, cross_val_score
from sklearn.feature_selection import SelectKBest, f_classif

# Numeric and categorical transformers
numeric_transformer = Pipeline(steps=[
 ('scaler', StandardScaler())
])

categorical_transformer = Pipeline(steps=[
 ('encoder', OneHotEncoder(handle_unknown='ignore'))
])

# Column transformer
preprocessor = ColumnTransformer(transformers=[
 ('num', numeric_transformer, numeric_features),
 ('cat', categorical_transformer, categorical_features)
])

# Full pipeline with feature selection + model
pipeline = Pipeline(steps=[
 ('preprocessor', preprocessor),
 ('selector', SelectKBest(score_func=f_classif, k=15)),
 ('classifier', GradientBoostingClassifier())
])

Cross-Validation & Model Comparison

Models were compared using Stratified K-Fold cross-validation (k=5) to account for class imbalance and reduce variance in performance estimates.

Model	CV AUC (mean ± std)
Logistic Regression	0.81 ± 0.03
Random Forest	0.87 ± 0.02
Gradient Boosting	0.89 ± 0.02

1
2
3


cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=42)
scores = cross_val_score(pipeline, X, y, cv=cv, scoring='roc_auc')
print(f"AUC: {scores.mean():.3f} ± {scores.std():.3f}")

Feature Importance

After fitting the final Gradient Boosting model, feature importances were extracted and visualized using matplotlib. The top features aligned well with domain knowledge, providing a sanity check on the model.

Key Takeaways

Pipeline design matters: keeping preprocessing inside a Pipeline object prevents data leakage during cross-validation, a subtle but critical mistake to avoid.
Model selection: Gradient Boosting outperformed simpler models, but the gain wasn’t free — it required careful regularization to avoid overfitting on the training set.
Feature selection: Using SelectKBest inside the pipeline reduced noise features and slightly improved generalization.

Tools Used

Full code available in: BU_OMDS_SU25_DX699B_RW

About

Thu, 01 Jan 2026 00:00:00 +0000

Who I Am

I’m Rafael Widjajahakim — a data scientist in training with over a decade of hands-on work in biomedical research and pharmaceutical market research. I’m currently completing my Master of Science in Data Science at Boston University while bridging the gap between clinical domain expertise and modern machine learning.

My path is a bit unconventional: I started in wet-lab research, moved into data-heavy clinical trial work, pivoted to pharma secondary research and analytics, and now I’m focused on building end-to-end data science skills that can make a real difference in health outcomes.

Experience & Background

Biomedical Research — ~10 years of experience across ophthalmology clinical trials (AREDS), lipid research, and real-world evidence (RWE) studies
Pharma Market Research — Line-of-therapy analyses, executive summaries, and dashboard reporting for pharmaceutical clients
Data Science Internship (6 months) — AWS EC2, Snowflake, Python-based data cleaning and pipeline work in a production environment
Teaching Assistantship / Coursework — Graduate-level modules in statistical modeling, machine learning, and data engineering

Technical Skills

Category	Tools
Languages	Python, R, SQL, SAS
ML & Modeling	scikit-learn, pandas, GridSearchCV, cross-validation
Stats	Biostatistics, hypothesis testing, regression, survival analysis
Visualization	Tableau, PowerBI, ggplot2, matplotlib, seaborn, Adobe Illustrator
Cloud & Data	AWS EC2, Snowflake, Git, Jupyter, VS Code
Formats & Pipelines	ETL workflows, feature engineering, model evaluation

Education

Boston University — MS in Data Science (in progress)
Relevant coursework: Machine Learning, Biostatistics, Statistical Programming, Data Engineering

What I’m Working On

Building a stronger portfolio of end-to-end ML projects in Python
Deepening expertise in gradient boosting, hyperparameter optimization (Optuna), and NLP
Exploring real-world healthcare datasets for predictive modeling
Improving data storytelling through clean, precise visualizations

Beyond Data

I’m a photography and videography hobbyist with a soft spot for Fuji cameras and drone footage of nature. I enjoy hiking, specialty coffee, and the ongoing challenge of getting through my ever-growing backlog of unedited RAW files.

Feel free to reach out — rafaelwh@bu.edu · LinkedIn · GitHub

Visualizing abstract formula for research paper

Wed, 02 Dec 2020 00:00:00 +0000

Line

tl;dr

Basically, AREDS clinical trial grading data was found to be flip-floping for our analysis purposes. We defined a certain criteria for our research that is unique but could cause confusion due to the complexity of the formula or uncommon technique in this dataset.

To address this, I created a simple visualization to help explain the flip-flop issue, and the result using our novel code.

Another good thing about this visualization is that, the code to make this graph can easily & quickly be modified depending if reviewers/leaders want to see a different version.

The original visualization was created using R + ggplot2.

BACKGROUND STORY

AREDS 1 overview

AREDS was a multiyear longitudinal clinical trial on age-related macular degeneration, looking to see the effect of supplements on the progression of the disease. During the first AREDS trial, the grading system utilized 9 steps plus progression to the advanced AMD.

Many papers have been published and explained the trial design and results. So I won’t go into details here.

–

Research Objective

While many studies now used the simplified scale, we used the full scale to create a more granular analysis on the time-based progression of the disease.

Severity scale 1-8: was for the non-advanced AMD grade.
Severity scale 9: was for the advanced AMD grade.
modified severity scale 9, to include 10, 11, 12: was our modification to include GA and NV as part of the continuous scale.

–

Story

Since we analyzed it in very granular level, we hit an issue where the count of advanced AMD was different between timepoints.
Upon further investigation, we found that there were many cases where an eye was graded as progressed to advanced stage but then went back to non-advanced stage.
Since we wanted to analyze this progression like survival analysis, we had to make sure that a progressed eye should not go back to non-advanced stage.
- In some cases, the advanced grade only occured once in the middle of the study period and then back to very early stage (scale <6).
- In other cases, the time when the eye progressed to advanced stage occurred multiple times.
So, we developed a criteria to only select the eye as progressed eye, when the grade or status had passed the progression status/scale, at least at 2 time points in a sequence.
Then we realized the system was quite abstract.

So, I developed a visualization to show the progression and the flip flopping issue.

–

GRAPH PREVIEW

Understanding the graph:

The graph shows the typical grading for the subjects in AREDS 1 trial.
- The usual grade was taken after baseline visit was within 2 years.
- Different subjects had different total follow-up time.
The main issue is represented on ID 5, 6, and 7 (eye ID, not the same as the AREDS ID, the replacement is just for better visualization).
- Eye went to advanced stage on 2nd grading (scale 11), but went back to low (scale 4) on the 3rd visit. It even reached healthy status (scale 1) on the 5th visit.
- We decided that on these cases, the progression could be due to some human errors or study design error (blinded graders).
- These cases were the reason we developed a different criteria in defining when an eye was progressed to advanced stage.
The next few cases are representing how we selected the time of the progression.
- Eye 11 is a good example on how we skip the first progression (scale 11) as the following grades were back to non-advanced stage.
- When the eye finally had advanced stage grade back to back, we defined that this was a true progression.
- But instead of selecting the first grade, we chose the second timepoint for the true termination point.

Other version

CONCLUSION

Snippet for the R code is in my github.

The code originally was developed for our internal meeting where we wanted to view and compare individual ID/grades when we came across some issues. So, it was originally created so that we could quickly review each id and the code has more features such as multiple progression criteria.

We simplified the visualization a bit for the publication purpose.

CITATION

Seddon JM, Widjajahakim R, Rosner B. Rare and Common Genetic Variants, Smoking, and Body Mass Index: Progression and Earlier Age of Developing Advanced Age-Related Macular Degeneration. Invest Ophthalmol Vis Sci. 2020;61(14):32. doi:10.1167/iovs.61.14.32

Eye Center Logo 2021

Tue, 01 Dec 2020 00:00:00 +0000

THE STORY

When I was working at UMass Medical School, the Ophthalmology Department was looking to update the department logo. Somehow I was having design fever again and wanted to dust off my design skill. I have been involved in my lab’s website and knew many people in the department, so why not.

DESIGN IDEA / THINKING

ICON / BASE DESIGN

Our lab (and many other research-side teams) has several genetic projects. I was looking into how to merge / mix / assimilate something that is perceived as genetic with something that is like an eye.

GENETIC

Few ideas I had about some visuals that are common in genetic studies are:

Blots (like western blot, Southern blot, northwestern blot, etc.)
Marking of mutations in a chromosome
Exon splices

EYE

I guess there is not much variety or there is no need for an out-of-the-box idea for an eye

A full eye, anterior view
Sagittal view of eyeball
The iris

COLOR

I had known about Pantone way before, but Pantone color of the year somehow became a big news recently and it felt the color had become a trendy influencers topic. The color of the 2021 was somehow a two-colors instead of one.

There some info on the meaning of each color, and I thought they could fit to the department vision or current situation. So I decided to just use these 2 colors.

CREATION

I came up with three different styles.

COMMON AMONG 3

I couldn’t come up with three totally different ideas; so, these 3 have very similar workflows:

I used Adobe Illustrator circular pattern
- First, I created a custom brush, based on some pattern of genetic icons I came up earlier.
- Then, I created a circle line vector.
- Finally, I applied the line using custom brush instead of a solid line.
The 2021 text is the same on all three
- The 20 uses the gray color to mark both the unchanging part of the year (the two thousand), as well as that we just passed the year 2020, and the 2020 being marked gray can mean as the off.
- The 21 uses the yellow (Illuminating) to mark the new year, and maybe the new meaning (we were just starting to get used to COVID19 life anyway).

ART 1

The eye part is based on the iris, while the line is more like line-chart but I can say it is like exon splices too. There is a flare on the inner lower left, just mimicking how eye icon has flare to show the transparency.

ART 2

The eye base is technically a full front eye. The original logo of UMass Eye Center used this style of eye as well (and many other eye clinics). The line part was inspired a bit from the Southern blot, but technically it was created using R as a stacked bar graph.

ART 3

The idea is similar to Art 1, but there is more flair added. The line part was like chromosome even though it was created with random ellipse tools.

SELECTION

In the end I selected Art 2 since I like the stacked bar graph line. Art 1 is good too, but it doesn’t really show like an eye. Art 3 is just too busy.

IN THE END

The deparment had a form to submit the designs. I even created a report form to explain my thinking:

But the decision makers in the end decided not to change their old logo at all 😓.

Readme for Excel file

Sat, 22 Aug 2020 00:00:00 +0000

BACKGROUND STORY

While data analytic work now have plenty of platform and tools, many business are still using Excel as their main program/software/app.

Some of my works used Excel for a simple dashboard or even a work report. Sometimes I saw some Excel dashboard that is quite complicated and I have seen my colleagues were confused on how to use it.

In other cases, we received different datasets in Excel files. They could be a new data or just a transformed version of the larger data. But sometimes, we received the files almost every day due to the nature of the project. This has caused some confusions on what the data is about, or worse, the version of the revision (when the filename itself does not tell you about it).

SOLUTION

We tried to duplicate how a good data science project will have a README file. But for Excel, we can always have one tab for this readme, while the other tab is the content of the file.

Links

Mon, 01 Jan 0001 00:00:00 +0000

To use this feature, add links section to frontmatter.

This page’s frontmatter:

1
2
3
4
5
6
7
8
9


links:
 - title: GitHub
 description: GitHub is the world's largest software development platform.
 website: https://github.com
 image: https://github.githubassets.com/images/modules/logos_page/GitHub-Mark.png
 - title: TypeScript
 description: TypeScript is a typed superset of JavaScript that compiles to plain JavaScript.
 website: https://www.typescriptlang.org
 image: ts-logo-128.jpg

image field accepts both local and external images.

Search

Mon, 01 Jan 0001 00:00:00 +0000

Rafael Widjajahakim

Hello World of Website Building

BACKGROUND STORY

A MORE ADVANCED PATH (why run before walk)

NOW

MORE HEADACHE

ML Pipeline: Predicting with Scikit-Learn

Overview

Problem & Dataset

Pipeline Design

Cross-Validation & Model Comparison

Feature Importance

Key Takeaways

Tools Used

About

Who I Am

Experience & Background

Technical Skills

Education

What I’m Working On

Beyond Data

Archives

Visualizing abstract formula for research paper

tl;dr

BACKGROUND STORY

AREDS 1 overview

Research Objective

Story

GRAPH PREVIEW

Understanding the graph:

The next few cases are representing how we selected the time of the progression.

Other version

CONCLUSION

CITATION

Eye Center Logo 2021

THE STORY

DESIGN IDEA / THINKING

ICON / BASE DESIGN

GENETIC

EYE

COLOR

CREATION

COMMON AMONG 3

ART 1

ART 2

ART 3

SELECTION

IN THE END

Readme for Excel file

BACKGROUND STORY

SOLUTION

Links

Search