AI Engineer in Bandung, Indonesia. I build LLM systems whose answers can be checked: every score cites its evidence and says how sure it is.
- LLM-as-a-judge at Indonesia's FDA (BPOM), 2025–2026. Scores senior-leadership selection papers on five rubric indicators, quotes the passage behind each score, and re-runs itself 5–7 times so the spread becomes a confidence value. Grounded with LightRAG on PostgreSQL/pgvector and Neo4j (RAGAS faithfulness 0.80–1.00). Served Qwen3 with vLLM behind LiteLLM on an on-prem GPU server, later Azure OpenAI.
- RAG for civil-service competency mapping. 461 regulation chunks, 94-query golden set, Hit Rate@3 78.7%, MRR 0.55, 100% JSON schema compliance. Paper accepted at ICAITech 2026.
- Agrimate, smart farming. TensorFlow Lite models and IQR-based sensor cleaning on edge devices over LoRa. Finalist, InnoServe Awards 2024 (Taiwan).
The BPOM code and documents are internal, so they are not public. Case studies with diagrams are on erwinhafiz.me and merita.
Python · SQL · LangChain · LangGraph · LightRAG · RAGAS · vLLM · LiteLLM · Azure OpenAI · FastAPI · Docker · PostgreSQL/pgvector · Neo4j · MLflow · GitHub Actions · Prometheus · Grafana
erwinhafiz.me · LinkedIn · erwinhafizt@gmail.com
Open to AI/ML and LLM Engineer roles (remote or Indonesia) and LLM evaluation contracts.