I build production AI systems — LLM pipelines, ML models, and the full-stack applications around them.
My degree is Mechanical Engineering, focused on process safety and reliability for oil & gas. That pairing is deliberate: most AI tooling for industrial work is built by people who have never sat through an FMEA. I've done both sides.
📍 Bekasi, Indonesia (GMT+7) · open to remote roles 🔗 Portfolio · LinkedIn · Hugging Face
AntiDeadline.ai — closed beta · Next.js 16 · TypeScript · Supabase A document-analysis SaaS that turns raw PDF/DOCX/XLSX files into auditable root-cause-analysis reports (5 Why, Fishbone/Ishikawa, full RCA). Runs a three-stage multi-model pipeline — parsing → evidence extraction → report generation — with strict schema validation at every boundary, JSON-repair retries, atomic credit transactions, and row-level security. Every report separates verified evidence from AI assumptions from missing information, so nothing reads as a confident guess.
prompt.aja — live · TypeScript Visual AI generation for Indonesian creators and small online sellers: text-to-image and text-to-video with 16 Indonesian-language prompt presets and platform-correct aspect ratios.
AI-automated field inspection reporting — serverless · Telegram Bot API Digitizes Management Walkthrough reports for oil & gas field operations. Operators send free text, photos, or voice notes; the bot transcribes audio, extracts structured data, runs a stateful clarification dialog, and populates Google Sheets with evidence archived to Drive.
| Project | What it does |
|---|---|
| aloha-ml-surrogate | ML surrogates replacing ALOHA dispersion simulations for real-time hazard-zone prediction. Trained on 1,215 full-factorial scenarios; cut held-out Red-Zone error to 1.82 m MAE against 22.25 m for the linear baseline — plus an out-of-distribution analysis mapping where the surrogate stops being trustworthy. |
| aloha-hazard-dashboard | Dockerized Flask inference app for those models, deployed on Hugging Face Spaces. |
| bi-dashboard-zona-bahaya | Plotly Dash BI layer — descriptive statistics, correlation analysis, model comparison, robustness testing. |
| ai4i-dashboard | CNC failure-mode classification on the AI4I 2020 dataset — Decision Tree, Random Forest, XGBoost and ANN benchmarked with SMOTE. |
- Promoting Evidence-Based Energy Transition: Statistical Analysis for Sustainability Policies in Indonesia — first author, SUSTINERE (Scopus / SINTA-2), accepted 2026. ARIMA forecasting, multiple linear regression, and a 10,000-iteration Monte Carlo simulation on Indonesia's renewable-energy targets.
- Comparative Analysis of HEART vs SLIM in Assessing Human Error Probability — first author, Motivection 7(1), 2025. doi.org/10.46574/motivection.v7i1.431
- Studi Kelayakan: Halide Perovskite sebagai Penanda Kimia dalam Industri Minyak dan Gas — co-author, Universitas Pertamina Press, 2026. Copyright registration No. 001135780.
Languages Python · TypeScript · SQL ML scikit-learn · TensorFlow/Keras · XGBoost · pandas · NumPy · SHAP LLM engineering multi-model pipelines · structured output & schema validation · prompt design · evaluation guardrails · RAG patterns App & data Next.js · React · FastAPI · Flask · PostgreSQL · Supabase · Drizzle · Docker Process safety FMEA/RPN · QRA · consequence modelling · HEART · SLIM · ALOHA