All Projects
The rest of the portfolio — earlier LLM projects, applied ML work, open-source NLP contributions, and a couple of things from further back. For the current highlights, see the Selected Projects on the homepage.
HR Attrition Analytics
A full data science workflow on IBM's HR attrition dataset.
- Approach: EDA, hypothesis testing (Welch's t-test, ANOVA), classification, regression, and clustering.
- Key Insight: OverTime turned out to be the single strongest attrition signal — employees working overtime leave at roughly 3x the rate of those who don't.
P1 · Aria — LLM Chatbot with Memory
A conversational AI with persistent memory, deployed live on Streamlit Cloud.
- Key Insight: The hard part of building a chatbot isn't the model call — it's state management.
P2 · DocChat — RAG Document Q&A
Upload any PDF and ask questions — retrieves relevant chunks first, then reasons on top. Deployed live on Streamlit Cloud.
- Key Insight: RAG is a retrieval problem wearing an AI costume — chunking and embedding quality matter more than the model.
P3 · AI Agent with Tools
A ReAct agent that searches the web, reasons across tool outputs, and decides what to do next.
- Key Insight: Building an agent is easy; controlling it is the real engineering.
P5 · Fintech LLM Analyst — Text-to-SQL
Ask a plain-English question about financial data, get a real answer backed by BigQuery SQL.
- Pipeline: The LLM generates the query, BigQuery runs it, and the LLM explains the result in plain English.
- Key Insight: Sits at the intersection of language, data, and decisions that affect real people.
Fintech Segmentation API
A production REST API serving a KMeans behavioral segmentation model trained on 278,932 transactions.
- Model: Classifies users into 4 segments across 31 behavioral features.
- Infra: Fully Dockerized with CI/CD via GitHub Actions.
Fintech Behavioral Segmentation
The research side of the segmentation work — clustering on fintech transaction data to identify distinct behavioral patterns.
- Features: Spending habits, merchant diversity, temporal patterns, and payment behavior.
Olist Analytics — dbt + BigQuery
An end-to-end data pipeline on the Olist Brazilian e-commerce dataset.
- Architecture: Modeled into a star schema with staging, intermediate, and mart layers.
- Downstream: The same BigQuery mart tables feed into the Fintech LLM Analyst above.
SEA Dataset Dataloader — SEACROWD
Open-source contribution to SEACROWD — standardized HuggingFace dataloaders for Southeast Asian language datasets.
- Impact: Part of the research effort that led to the paper at EMNLP 2024.
Face Recognition — Computer Vision
A real-time face recognition system built from scratch.
- Pipeline: Face detection, matching, and identity classification.
- Note: The project where I first realized making a computer see things was way more interesting than expected.
Twitter Sentiment Analysis — Football Fan Bases
Comparative sentiment analysis of Indonesian Twitter discourse around Real Madrid and FC Barcelona fan communities.
- Methods: Lexicon-based methods and ML classifiers on scraped Twitter data.
Stock Trading Robot
An automated stock trading agent built for a mathematical modeling course.
- Model: LSTM predicts stock prices up to ten days ahead and recommends buy, sell, or hold.