Machine Learning Projects

Machine Learning

Invoice Document to Text

An end-to-end deep learning project that converts invoice images directly into structured JSON data.

Built a custom encoder-decoder architecture using a Swin Transformer for visual feature extraction and a BART Decoder for autoregressive text generation. Developed the complete pipeline including dataset preparation, image preprocessing, tokenizer development, model training, evaluation, and inference using PyTorch. Used MLflow for experiment tracking and Google Vertex AI for GPU-based training workflows. The trained model was published to Hugging Face Hub for deployment-oriented inference.

PyTorchPyTorchSwin TransformerBARTMLflowMLflowVertex AIVertex AIHugging FaceHugging Face

Product Price Prediction using Gradient Boosting

A machine learning solution for predicting e-commerce product prices using multiple tree-based and distance-based regression models.

Applied data cleaning, exploratory data analysis, ordinal encoding, Min-Max normalization, and log-transformed regression targets. Benchmarked XGBoost, LightGBM, CatBoost, Random Forest, Extra Trees, and KNN using MAE, RMSE, MAPE, and R². XGBoost achieved the strongest overall performance.

XGBoostLightGBMCatBoostRandom ForestFeature Engineering

EasyOCR Under the Hood: CRAFT and CTC Recognition Analysis

Explored the internal architecture of EasyOCR by decomposing its complete OCR pipeline into modular experiments.

Investigated CRAFT for text detection, the VGG-based english_g2 recognizer, and multiple image preprocessing techniques affecting OCR performance. Compared Greedy Search, Beam Search, and Word Beam Search for CTC decoding while analyzing prediction probabilities and recognition outputs.

OCRCRAFTCTCComputer VisionBeam Search

TinyStories-1M Indonesian Fine-Tuning Experiment

An experimental project exploring fine-tuning of the TinyStories-1M language model using an Indonesian corpus.

Demonstrates the end-to-end workflow of training, evaluating, and analyzing an Indonesian language model using the Lyon28/Corpus-Indonesia dataset.

LLMFine-tuningPyTorchPyTorchHugging FaceHugging FaceIndonesian NLP