Apri Dwi Rachmadi
AI/ML engineer and Statistics graduate who ships ML systems end-to-end — from LLM evaluation and RAG to computer vision and production tabular models.
Projects
Selected work, with the receipts.
Systems I designed, trained, and built — with the metrics that matter.
Click any card to expand the full breakdown.
Experience
Where I've shipped.
Owned the LLM benchmarking codebase, scoring emerging models against proprietary datasets. Built an automated "LLM-as-a-Judge" pipeline reaching 84% alignment with human evaluators on RAG & agent use cases, and productionized tabular models (demand forecasting, loan-fraud and anomaly detection) with drift monitoring, automated retraining, and unit testing. I also favor explainable ML — the attendance-based anomaly-detection system doesn't just flag what's anomalous, it produces human-readable explanations of why each case was predicted as an anomaly.
Built an end-to-end automated cashier system (YOLO11 detection at 82% mAP, MobileNet classification at 99%, zero-shot segmentation for auto-annotation), a Llama + LangChain + ChromaDB RAG audit assistant, a PaddleOCR receipt-processing pipeline, and a collaborative-filtering training recommender — served via a Flask backend.
Part of the Data Science & Analytics division. Built a comprehensive Tableau dashboard for the KRL commuter line, analyzing schedule busyness and the number of operating stations and trains.
Graduated with Distinction (score 94.24/100). Studied ML foundations through deep learning and model deployment. Capstone UPCYCLE: a TensorFlow garbage classifier at 91% accuracy, deployed to Android via TensorFlow Lite.
GPA 3.73/4.00 (147 SKS). Thesis: TrOCR for vehicle license-plate recognition. Teaching assistant for Algorithms & Programming, and a Satria Data finalist three years running (2022–2024).
Awards & Honors
Tested in national competitions.
Data-science and ML competition results across three years, including three consecutive Satria Data finals.
LightGBM topic classification on political tweets in the preliminary round; in the finals, LDA topic modeling and RoBERTa sentiment analysis on discussion around the 2024 presidential election.
A YOLOv8 detection model (99.5% mAP) plus EfficientNetV2B0 number classification (99.84% F1) to automatically count votes on SIREKAP forms.
Data preprocessing, multi-table joins, feature engineering, and EDA to build an ML model for credit-scoring classification.
Preprocessing, visualization, and a LightGBM classifier predicting network-attack type, reaching a 99.58% F1-score.
A TrOCR model for licence-plate recognition in the preliminary round; in the finals, a Siamese BERT to detect disharmony in capital & taxation laws.
Damage-grade classification for earthquakes and image classification for early fire detection using deep learning.
Analysis of patient-return status for BPJS Health participants and public-opinion analysis of BPJS services on Twitter using IndoBERT and BERTopic.
About
I turn statistics & research into working ML systems.
Statistics graduate of Universitas Diponegoro and a Bangkit Machine Learning Distinction graduate — I build models end-to-end across LLMs, computer vision, NLP, and tabular data, with a strong analytical foundation honed through national data competitions.
Languages
ML / Deep Learning
LLMs & NLP
Computer Vision
MLOps & Serving
Data & Analytics
Contact
Let's build intelligent systems together.
Open to AI / ML Engineer and Data Scientist roles. The fastest way to reach me is email.
rachmadiapri@gmail.com