Projects by Muhammad Faza Pinaringgi: LLM scoring, Gemini agents, medical imaging research, and web products.
Four live products I build and run as the only engineer.
LiveJul 2026 - present
A paid language pre-test platform where Claude scores Writing and Whisper transcribes Speaking, checked against real IELTS results.
Next.js 15 TypeScript Supabase/PostgreSQL Claude API +4 −0.15 band −0.15 band , Details for: Writing-scorer bias Writing-scorer bias Mean bias on 46 band-scored reference essays; 82% within half a band (band 4-5.5). Writing-scorer bias
Read case study LiveAug 2026 - present
I redesigned the whole site, then built what runs the business behind it: a member area, an admin panel, lead scoring, and online payments.
Laravel PHP MySQL Midtrans +1 Live2026
A full redesign of a job portal for Indonesians who want to work abroad.
Laravel Blade Bootstrap JavaScript Live2026
A full redesign that turned a site for visa applicants into a clear B2B site for Australian employers.
Laravel Blade CSS JavaScript Research papers, course projects, and prototypes, each with its stack.
All AI/ML Research Web 9 projects
Project Oct 2026
A 10-day prototype for a team in an internal corporate AI competition. The team won 1st place.
React 19 TypeScript Gemini function calling Vercel serverless 14 14 , Details for: Query tools for the analyst agent Query tools for the analyst agent Every number in an answer is computed by code, not generated. Query tools for the analyst agent
Read case study In progress 2025 - present
Master's thesis: three different teachers trained on 112,120 chest X-rays, distilled into one lightweight student.
PyTorch DenseNet-121 ResNet-50 RAD-DINO ViT +1 0.840 0.840 , Details for: Best teacher mAUC Best teacher mAUC RAD-DINO ViT on 112,120 NIH chest X-rays; CheXNet reports 0.841. Best teacher mAUC
Read case study Published 2025
LSTM rainfall forecasting feeding an ANUGA flood simulation, rendered as an interactive 3D map. Published at ICCSCI 2025.
0.998 0.998 , Details for: Rainfall LSTM R² Rainfall LSTM R² MAPE 3.3%; the forecast feeds an ANUGA flood simulation. Rainfall LSTM R²
Read case study Published 2024
Which balancing technique helps when only 0.172% of 284,807 transactions are fraud? Published at ICCSCI 2024, presented orally.
Python scikit-learn XGBoost imbalanced-learn 92.4 % 92.4 % , Details for: Best fraud-class F1 Best fraud-class F1 XGBoost with random oversampling; 284,807 transactions, 0.172% fraud. Best fraud-class F1
Read case study Project 2024
Classical models versus CNNs on PlantVillage leaves, served through a web app with a low-confidence guard.
TensorFlow MobileNet ResNet scikit-learn +3 0.993 0.993 , Details for: MobileNet macro-F1 MobileNet macro-F1 On 3,943 test leaves; Random Forest 0.90 and SVM 0.88 on hand-made features. MobileNet macro-F1
Read case study Project 2025
A general-purpose model beat the domain-specific one on 3-class financial sentiment.
Hugging Face Transformers RoBERTa FinBERT BiLSTM 82.8 % 82.8 % , Details for: RoBERTa accuracy RoBERTa accuracy Weighted F1 0.835 on 877 test headlines; FinBERT 81.0%, BiLSTM 59.1%. RoBERTa accuracy
Read case study Project 2026
Three modern vision backbones compared on 7 imbalanced classes, judged by macro-F1.
PyTorch timm ConvNeXt V2-Tiny DeiT III-Small +1 0.728 0.728 , Details for: ConvNeXt V2 macro-F1 ConvNeXt V2 macro-F1 Up from 0.635; AUROC 0.957 on 7 imbalanced classes. ConvNeXt V2 macro-F1
Read case study Project 2024
Type your symptoms and get a likely disease with its description, precautions, medications, diet, and workout. A Flask app on top of five compared classifiers.
Python scikit-learn pandas Flask +1 Project 2023
Book a tutor by level, subject, and method. Designed in Figma, built in HTML, CSS, and JavaScript, with a MySQL schema.