An on-device Arabic script lens: photograph a letter and a MobileNetV2 classifier returns the Hijaiyah class, running inside an Android app through TensorFlow Lite.
45,308
training images
dataset_v5 · 29 classes
98.41%
validation accuracy
88.57%
UAT score
Python
TensorFlow
MobileNetV2
TFLite
Kotlin
Android
Built as a scientific project with a measurable goal: assemble a 45,308-image dataset across 29 classes, fine-tune MobileNetV2 with transfer learning, and ship the model onto a phone.
The model was evaluated both numerically (validation accuracy, per-class precision/recall, confusion matrix) and with user acceptance testing on the full mobile flow.
DISPLAY
+04 INSTRUMENTS
tadabburlens / splash-screen
01 / 10
INSTRUMENTATION 04 VIEWS
◂ SCROLL ▸ · CLICK TO REVIEW
02AUTOMATION · AI INTEGRATION
CASE 02 / 03
YoLiYo Mentai
Full-Stack Developer Team project · Pengelolaan Proyek Perangkat Lunak · 2026
An ordering system for a real MSME client: customers order through a Telegram chatbot (n8n + AI agent), payments run through Midtrans, and the owner runs the business from a React dashboard backed by Supabase.
100%
UAT
accepted by the business owner
React
Supabase
n8n
Telegram Bot
Midtrans
AI Agent
Group coursework required a real client — the client here is a family-owned food MSME. Full-stack development of the system (database, chatbot, dashboard, UI/UX, user testing) was carried out by me; teammates handled project management artefacts (BRD/PRD) and product direction.
Published with the team's permission. Integration depth is described from the built system itself.
DISPLAY
+04 INSTRUMENTS
yoliyo-mentai / login-page
yoliyo-mentai
01 / 15
INSTRUMENTATION 04 VIEWS
◂ SCROLL ▸ · CLICK TO REVIEW
03OCR · NLP · MOBILE PIPELINE
CASE 03 / 03
TashkiFlow
Thesis Researcher & Developer Undergraduate thesis research · 2026
An Arabic OCR and diacritization pipeline: scan a page, a hybrid OCR engine (PaddleOCR + Kraken, arbitration-merged) extracts the text, a sequence model restores the harakat, and an Android app presents Arabic + diacritics + Latin transliteration.
13.79%
average CER
character error rate
23.15%
average WER
25.75%
average DER
LIMITATION Known limitation: classical handwritten Arabic remains difficult for this pipeline — character error rate rises to 40.52% on that subset. The results above are not a claim of flawless accuracy.
Python
PaddleOCR
Kraken
Diacritization
FastAPI
Android
Measured per stage rather than with one accuracy figure, because each stage fails differently — character, word and diacritic error rates are reported separately.
Iterated through 20+ pipeline builds (v79 → v80) with per-stage evaluation before the final architecture was fixed.
DISPLAY
+04 INSTRUMENTS
tashkiflow / splash-screen
01 / 10
INSTRUMENTATION 04 VIEWS
◂ SCROLL ▸ · CLICK TO REVIEW
LOG / 002
Research context
Fresh graduate in Informatics Engineering from Gunadarma University, focused on artificial intelligence and machine learning — specifically computer vision and natural language processing.
Undergraduate research sits in Arabic OCR and diacritization: an area where models are still fragile, and where honest measurement matters more than a good demo.
TUTOR & LABORATORY ASSISTANT — LePKom
Java
Database / DBMS
Golang
Supporting coursework and lab sessions across these subjects.
The through-line: taking AI/ML experimentation and giving it a shape someone can actually use.
.n-ZWwXwX_2wBClr.webp)
.png.CZIgT8IG_Z177WQs.webp)
.Pq6I7F5O_hXfjW.webp)
.png.kl4XLLbr_Z1POioz.webp)
.CY-X_R7b_Z1vbJzC.webp)
).png.DoEzdgfH_SrKzK.webp)
01 / 10 
