− README overstated what the deployed model could do
− no recall analysis — accuracy only
+ accuracy 84 → 83%, a deliberate trade
+ hate-speech recall 0.14 → 0.18
AI · ML Engineer — Bogor, Indonesia
I went back and madeevery numberhonest.
BankruptWatch — the old score hid that the model caught none of the real bankruptcies. Honest recall: 57%, at 40.3% precision.
About
- PublishedSentiment Classification of Hate Speech Against Islam on Twitter Platform Using Multinomial Naïve Bayes. Journal of Informatics and Science Media (JISMEDIA), Vol. 1, No. 2, pp. 34–41, December 2024. ISSN 3064-1942.
- StudiedBSc Computer Science — GPA 3.50 / 4.00, UIN Maulana Malik Ibrahim Malang
- BelieveWhether a reported result is true matters as much as whether the model ships.
The reworks · 6
− "87% accuracy" — trained on 798 masks mixed into 780 real images
+ 69% val accuracy on masks-excluded data
+ per-class recall: benign .91 / malignant .47 / normal .14
− "92% accuracy with SMOTE" — matched nothing in the notebook
− SMOTE computed but never used in training
+ Random Forest recall 56.8% vs 0% baseline
+ 40.3% precision, stated with the trade-off
− every page embedded a dead localhost:8080 iframe
+ self-contained Leaflet + vanilla JS
+ live on laundrygis.aljuhaeda.com
− reported only the flattering accuracy
+ accuracy + macro-F1 for both models
+ TF-IDF 82.8/.80 · IndoBERT 89.7/.87
+ fixed an empty-input crash
− 4 crash bugs since coursework submission
− NetBeans run config pointed at a deleted class
+ compiles clean on JDK 23
+ builds and runs, verified end to end
Also shipped · 2
− MusliMalang (one city) + salatwebapp (1 of 5 prayers)
− both stalled
+ one Flutter app: all cities, all 5 prayers
+ themed reflections + 114-surah reader
+ live, real tenant & finance data
+ Supabase RLS enforced per staff role
Open to ML roles
If you want someone who checks the number before the demo — let's talk.
Email me →