Danish Rafie Inference ’26 Curriculum Vitae Get in touch
Inference ’26Surabaya, IDPortfolio of Danish Rafie Ekaputra

The
Signal
Edition

Turning messy data into models people can use. Research, analytics, and teaching at ITS.

5 papers5 national results4 courses taught11 countries mapped
I

Overview

A data science undergraduate who ships models, then makes them explain themselves.

OverviewI

Who’s behind the numbers

Danish Rafie Ekaputra — Data Science undergraduate at Institut Teknologi Sepuluh Nopember, Surabaya. GPA 3.52/4.00, class of 2023.

I care about the whole path, from cleaning the data to checking that a model still works on data it has never seen.

Research

Five papers in 2026, one published and four under review, from sign language recognition to battery X-ray inspection.

Industry

Power BI dashboards for Telkomsel loan analysis, expansion analysis for a national F&B franchise, and flight-control code for a competition UAV.

Teaching

Teaching assistant and instructor for four ITS courses since 2025, from data wrangling to machine learning.

Languages

Indonesian (native) · English (professional working proficiency)

3.52
GPA / 4.00
5
Papers
5
Awards & results
4
Courses taught
5028
Samples annotated
II

Research

Two papers ask where a photograph was taken. One asks what an advertisement is hiding. One counts the plates inside a battery. One tests sign recognition on signers the model has never seen.

ResearchII

Evidence, made explainable

Published2026Jurnal Statistika dan Komputasi

Multimodal Deep Learning for Online Gambling Promotion Detection on Indonesian Social Media

A multimodal fusion approach that reads image and text together to catch gambling promotions on Indonesian social media — plus the first publicly available multimodal dataset for the task: 5,028 annotated samples from Facebook, TikTok, and X.

Six fusion strategies evaluated across vision backbones (DINOv3, CLIP-ViT-L/14, SigLIP 2) and text encoders (Indonesian BERT, Qwen3-Embedding, Multilingual E5-large).

97.18%
Best F1-score
5028
Annotated samples
6
Fusion strategies
Manuscript under review2026

Signer-Independent SIBI Recognition: A Benchmark of Skeleton and Pretrained Models

A benchmark of five sign recognition models on 114 isolated signs of the Indonesian Sign System — affixes, numbers, letters, and words — recorded by 22 signers. Every model is tested on signers it never saw in training.

Earlier SIBI results were measured on signers seen in training, which overstates accuracy: the BiLSTM of that work falls from 97.29% to 85.72% on new signers. A fine-tuned Uni-Sign reaches 95.16% on the same new signers, and loses 5.6 points when the face landmarks are removed.

95.16%
Accuracy, new signers
22
Signers
114
Sign classes
Manuscript under review2026

Adapting Point-Based Crowd Counting to Power Battery Detection

Crowd-counting machinery pointed at battery X-rays: APGCC adapted to count the anode and cathode plates in the PBD5K benchmark and locate every endpoint, with a density-map counting head added.

Structure-aware decoding uses what a battery guarantees — plates arrive in a periodic order, and there is always one more anode than cathode. Mean pair-number accuracy reaches 0.8689 on the 2,000-image test set, against 0.7976 for MDCNeXt rescored with the official evaluator.

Undergraduate thesis
0.8689
Mean PN-ACC
2,000
Test X-rays
3
Difficulty splits
Manuscript under review2026

GeoSEAN11K: Separating Text and Scene Evidence for Country-Level ASEAN Image Geolocation

11,000 street-level images across the 11 ASEAN countries, balanced at exactly 500 text-bearing and 500 no-text images per country — so reading a sign and recognising a streetscape are measured apart.

Text-bearing scenes go to a zero-shot Gemma 4 26B that transcribes the sign and names the country in one response; no-text scenes go to a fine-tuned MaxViT-Small. Pooled over all 11,000 images, the two routes are right 78.75% of the time.

78.75%
Pooled accuracy
81.95%
No-text, MaxViT-Small
11,000
Street-level images
Manuscript under review2026

GeoSEAN: Explainable Country-Level Image Geolocation for ASEAN Regions

An explainable country-level geolocation pipeline across 11 ASEAN countries: CLIP zero-shot classification feeding LightGBM and MLP classifiers, with the MLP taking the best result.

Visual evidence is audited rather than assumed — CLIP attention rollout, YOLO26 object detection, and the Energy-Based Pointing Game measure whether the model looks at what actually localises a country.

85.91%
Accuracy, MLP
85.91%
F1-score
11
ASEAN countries

What each method added

baseline → this work · same metric in each row, %
Gambling detectionF1 · text only → image + text fusion
+1.29 pts95.89 → 97.18
SIBI sign recognitionAccuracy, new signers · earlier BiLSTM → Uni-Sign
+9.44 pts85.72 → 95.16
Battery plate countingMean PN-ACC · MDCNeXt → adapted APGCC
+7.13 pts79.76 → 86.89
GeoSEAN11KCountry from sign text · OCR → Gemma 4 reading
+28.05 pts33.33 → 61.38
GeoSEANAccuracy · CLIP zero-shot → MLP on CLIP features
+23.46 pts62.45 → 85.91
III

Awards

Five national results, each one with a number attached to it.

AwardsIII

Ranked, repeatedly

2026
F.I.R.E Scholarship Awardee
MEXC Foundation & TRIV

One of only 8 recipients chosen from hundreds of applicants across 6 top Indonesian universities. Full one-semester tuition plus mentorship in blockchain and Web3 innovation.

1 of 8
2025
Finalist (Top 3), Data Quest
Data Science Indonesia

A machine learning classifier predicting customer term-deposit subscriptions, ranking first on the preliminary leaderboard and reaching the top 3 in the national final.

AUC 0.7982
2025
Finalist, Data Mining
Intelecta Cup, Telkom University

A regression model forecasting average temperature, finishing top 3 among national competitors.

MAPE 83.9
2025
Finalist, SSDS
Sebelas Maret Statistics Data Science

A regression model predicting daily rainfall totals, finishing top 6 among national competitors.

MSE 17.05
2024
Best Methodology Award
Technology Development, KRTI

Recognised for a systematic methodology across the team’s UAV design and engineering process — the documentation, not just the aircraft.

UAV / KRTI
IV

Experience

A consulting floor, an AI community, and a UAV research team that had to actually take off.

ExperienceIV

Where the work happened

Data Scientist Intern
2026
Enciety Business Consult · Surabaya

Built an interactive Power BI dashboard for Telkomsel loan analysis, surfacing loan performance against customer behaviour. Ran data-driven analysis behind Kebab Turki Baba Rafi’s expansion strategy — which markets, which growth.

Flight Controller Division
2023 — 2025
Bayucaraka UAV Research Team, ITS

Engineered and programmed the flight controller on top of the ArduPilot open-source platform, implementing custom control algorithms for the team’s competition aircraft.

Fig. 1 · Bayucaraka’s Soeroku87, KRTI 2024, Yogyakarta
Data Scientist Division
2024 — now
Avalon AI Community

An ML/DL development program covering causal inference, ensemble learning, time-series modelling, and advanced feature extraction for image and audio data.

V

Teaching

Four appointments in two years — the fastest way to learn something twice.

TeachingV

Four rooms at ITS

2026
Teaching Assistant
Data Mining

Python practicums across the full pipeline: preprocessing, dimensionality reduction, association rules, K-Means / K-Mode / K-Prototype, Decision Tree and Naïve Bayes, then evaluation for both classification and regression.

2026
Instructor
Machine Learning, SCC

Data cleaning (imputation, deduplication, format standardisation), EDA with visualisation, and fundamental algorithms — KNN and Decision Tree — through to tuning with Grid Search, Random Search, and Optuna.

2025
Teaching Assistant
Sampling & Data Wrangling

Applying sampling theory and wrangling technique in Python lab sessions — cleaning, transformation, and the exercises that make the difference stick.

2025
Teaching Assistant
Exploratory Data Visualization

Practical sessions in Power BI, R, and Python — Matplotlib, Seaborn, and Plotly — aimed at charts that answer a question instead of decorating one.

VI

Toolkit

Instruments I actually reach for — not a list of everything I have ever installed.

ToolkitVI

Instruments

Languages
04 items
PythonRC++SQL
ML & Deep Learning
12 items
Multimodal learningComputer visionNLPClassificationRegressionClusteringFeature engineeringHyperparameter tuningTime seriesExplainable AIPose estimationVision-language models
Libraries & frameworks
17 items
PyTorchTensorFlowKerasTransformersScikit-learnAutoGluonOptunaPandasNumPyMatplotlibSeabornPlotlyMLlibLightGBMOpenCVMediaPipeR Shiny
Tools & platforms
11 items
Git & GitHubPower BIPostgreSQLFigmaArduPilotKaggleHugging FaceGoogle ColabLaTeXLinuxROS 2
Chapter VII

Let’s put this
to work.

Open to research collaboration, data science internships, graduate study, and freelance analytics work. Fastest route is email.

danishrafie.ep@gmail.com View CV
LinkedInin/danishrafie Google ScholarProfile ORCID0009-0008-8070-3692 GitHub0xRafie Kaggledanzarchive Hugging Face0xRafie
© 2026 Danish Rafie EkaputraBack to cover ↑