Every build starts
with a question.

Loan analysis

Random Forest model accuracy

97.40%

#1 feature in Random Forest and XGBoost

Income

LogReg after log transform + scaling

+1.49 pp

Problem

Which customer attributes relate to personal-loan uptake, and how do classification approaches compare?

My work

Analyzed customer data and compared classification models, delivering a notebook, model artifacts and a written report.

Result

Random Forest recorded 97.40% accuracy. Tree-based feature importance ranked Income first in both Random Forest (56.34%) and XGBoost (39.72%). For Logistic Regression, applying a log1p transform with RobustScaler increased recorded accuracy from 88.72% to 90.21%, a +1.49 percentage-point improvement.

Makemore

parameters built and trained from scratch

10M+

character context window

256

Problem

Understanding a language model requires more than calling an API. I wanted to follow the mathematics from a bigram baseline to a working GPT.

My work

Built a character-level GPT from scratch using PyTorch, implementing attention, training and Indonesian poetry generation.

Result

Delivered a configurable GPT module with six Transformer blocks, six attention heads per block and a 256-character context window, plus training/inference scripts and generated poetry examples.

Dio the Chatbot

to understand Hans quickly

1 chat

Problem

A recruiter should find relevant evidence without reading every portfolio page.

My work

A conversational RAG guide that answers questions across Hans’s projects, experience and source material.

Result

An interactive guide that lets visitors ask about Hans’s work and get source-backed answers without reading every portfolio page.

Monelytics

analysis coverage

4 banks

team modeling scope

9 approaches

Problem

Comparing forecasting approaches requires a consistent data pipeline and evaluation rather than a single prediction.

My work

Led EDA, preprocessing, feature engineering and LSTM implementation for a team stock-analysis application.

Result

Compared nine approaches for four Indonesian banks. The team delivered five modeling notebooks, four bank-specific LSTM files and a Streamlit application with RMSE/MAE comparisons.

Whisper fine-tuning

relative WER reduction

25.62%

German evaluation utterances

16,082

Problem

How does adapting a small speech model to a specific language change recognition?

My work

Adapted Whisper-Tiny to German speech, with reproducible training, evaluation and published model weights.

Result

The V2 notebook records word error rate falling from 43.49% to 32.35% across 16,082 German utterances: 11.14 percentage points, or 25.62% relative reduction. Published German and French model variants with training logs.

ResNet on CIFAR-10

reported test accuracy

87.92%

gain across configurations

~12.92 pp

Problem

Understanding residual networks means testing how architecture and augmentation affect image classification.

My work

Built a configurable image classifier and compared architecture and augmentation choices on CIFAR-10.

Result

Reported 87.92% test accuracy on 10-class CIFAR-10, approximately 12.92 percentage points above the earlier ~75% configuration. Delivered training, evaluation, checkpoint and resume workflows.

Narrative Nest

storyboard capacity

9 frames

generation configuration

8 steps

Problem

Filmmakers need a quick way to turn written scene ideas into storyboard frames.

My work

Led UI/UX, React frontend and AI integration for a shared-prompt storyboard creator.

Result

Delivered a nine-frame storyboard interface using 8-step SDXL-Lightning generation, with shared/per-frame prompts, editable shot notes and Firestore-backed generated-image records.

NeuroCraft

MNIST and Fashion-MNIST

2 datasets

network layer choices

5 controls

Problem

Beginners need to experiment with neural-network layers without building an entire training application.

My work

Built a visual interface for configuring, training and exporting small neural networks.

Result

Exposed two image datasets and five layer controls, backed by a Flask/PyTorch training API with CPU/CUDA selection and timestamped model export.

MNIST from scratch

recorded validation accuracy

90.08%

trainable parameters

7,960

Problem

I wanted to understand the operations that turn handwritten pixels into a prediction.

My work

Implemented the mathematics of a digit classifier, from initialization to manual backpropagation.

Result

The later PyTorch tensor notebook records 90.08% validation accuracy on 16,800 images with 7,960 trainable parameters. Delivered original NumPy and GPU-capable tensor implementations.

World Cup semifinal analysis

comparison scope

4 teams

cleaned analysis outputs

3 datasets

Problem

A semifinal preview needs to connect team workload, attacking performance and player statistics into a readable comparison.

My work

Translated match and player statistics into a comparative semifinal-preview analysis.

Result

Compared four national teams and delivered three notebooks, three cleaned datasets, two exported charts and an editable presentation.

LookSharp

AI-assisted build · Codex + Cursor

eyewear reference inputs

3 paths

default image upload limit

10 MB

Problem

Trying eyewear online needs a useful preview without losing the identity and context of the original portrait.

My work

Built a multimodal eyewear try-on app that connects image generation, validated uploads and a persistent gallery.

Result

Delivered three eyewear reference paths: catalog selection, custom upload and image URL. Added a 10 MB default upload cap, generation timing and an IndexedDB gallery with save/delete operations.

Micrograd, rebuilt

Problem

Automatic differentiation can feel like a black box when a framework does all the work.

My work

Rebuilding the Micrograd library from scratch to understand the underlying math and operations of a simple deep learning library. Implemented forward and backward propagation, gradient descent, and backpropagation using only pure Python Syntax.

Result

A pure-Python implementation of forward propagation, backpropagation, and gradient descent.

IDRXY

AI-assisted build · Codex + Cursor

index + reference currencies

9 + 12

snapshot cache duration

60 sec

Problem

Rupiah strength is often judged only against the US dollar. A basket-based view, similar in approach to DXY, helps show how the Rupiah moves against multiple currencies.

My work

Implemented a weighted Rupiah index dashboard with timestamped snapshots and explicit handling of missing chart data.

Result

Delivered a nine-currency weighted index and 12 additional FX references, with Supabase snapshot retrieval, 1,000-row pagination and a 60-second data cache.

Krakatoa

AI-assisted build · Codex + Cursor

interactive 3D atlas

9 chapters

reviewed automated suite

13 tests

Problem

Volcanic history, terrain and monitoring data are difficult to understand when scattered across disconnected sources.

My work

Built an interactive 3D field guide with public-data feeds, historical terrain states and resilient parsers.

Result

Delivered nine chapters, five historical states and six wind levels. Two APIs integrate three public-data sources; 13 automated tests and the reviewed test/build CI run passed.