Portfolio

I’m a Data Scientist and AI Engineer. Most of my day job is turning messy operational problems into something a person can act on: a dashboard, a chatbot, a camera pointed at a powder sample. On the side I work on Indonesian language models and small tools that scratch my own itch.

Work & Impact

  1. OneAnalytics Internal analytics and dashboarding platform A single place for the whole company's business metrics, replacing a request-and-wait reporting process. 50+ daily users, manager level and above PythonPlotly DashSQLBigQuery
  2. Self-service analytics chatbot Data-grounded internal assistant An LLM that answers business questions directly against company data, grounded hard enough to be trusted. Time-to-insight: 25 hours → 3 minutes LLMRAGPythonSQL
  3. Automatic dustiness powder QC Computer vision for quality control Video analysis that measures the dust plume from falling powder, turning a subjective QC judgment into a number. 100+ samples measured daily Computer VisionSegmentationPython
  4. Receipt & invoice recognition Hybrid OCR and LLM document parsing Automated extraction from receipts and invoices, built to run on edge devices. 92% accuracy on internal benchmarks OCRLLMPythonEdge deployment
  5. Data pipeline reliability 1st Place, Quality Control Circle Took a critical pipeline from recurring SLA breaches to full compliance, and won the group-level QCC. 100% SLA compliance on ~27K runs per week Data EngineeringMonitoringProcess design

Research

  1. Anak Baik A low-cost approach to curating Indonesian ethical and unethical instructions An Indonesian dataset for teaching language models to tell ethical instructions from unethical ones. Published at SEALP 2025. Peer-reviewed, pp. 52-62 Language ModelingDataset CreationLoRA
  2. AI-generated image classification Fine-tuned residual networks for detecting synthetic images Detecting whether an image was generated by AI, using fine-tuning techniques on ResNet. Published in JTIIK. F1 score 0.963 Computer VisionResNetPyTorch

Open Source & Lab

  1. Marzuki-7B An Indonesian LLM, end to end Continued pretraining, SFT, and DPO alignment on top of Mistral-7B. Three checkpoints published. 3 models on HuggingFace PyTorchLoRADPOHuggingFace
  2. Herman Indonesian JSON-mode dataset An Indonesian adaptation of the Hermes function-calling dataset, for training models to emit reliable structured output. Published dataset on HuggingFace Dataset CreationSFTJSON Schema
  3. Artinya An LLM translation pipeline Translation that tries to preserve tone, nuance, and cultural meaning, then evaluates whether it actually did. LLMPythonStructured output
  4. Rec2Slide Video recording → PDF slides Reconstructs a slide deck as a PDF from a screen recording. Published on PyPI. pip install rec2slide PythonOpenCVPyPI
  5. Batch big data ETL DuckDB, Spark, Pandas, Postgres A dockerized ETL pipeline with star-schema dimensional modeling, built to compare processing engines. DuckDBSparkPostgresDocker
  6. Kuantil Quant finance modules Option pricing and quant primitives as small, tested Python modules. Pythonpytestuv
  7. Fanesis Facial network analysis Maps relationships inside a group from images, using face recognition and emotion detection. Computer VisionGraph AnalysisPython

Everything else lives on GitHub.