At Sigma Solusi Indonesia, I work as a Data Scientist building end-to-end ML solutions for supply chain optimization and intelligent automation. The work spans demand forecasting, conversational AI, LLM fine-tuning, and automated retraining — all serving a B2B FMCG distribution platform.
RAG Chatbot for B2B FMCG
Architected the foundational chatbot layer for the platform using a Dockerized RAG system built on ChromaDB and sentence-transformers. The core challenge was that customers use colloquial and regional product names that don't match official catalog entries — "mie goreng bungkusan merah" won't keyword-match any SKU. The solution uses semantic embeddings across three unified collections (products, FAQs, intents) to resolve queries regardless of naming variation. The system handles 35+ customer intent types through semantic similarity against example phrases — no hardcoded keyword matching.
LLM Fine-tuning for WhatsApp Deployment
Engineered a specialized FAQ model for WhatsApp deployment using knowledge distillation and LoRA fine-tuning — distilling a Llama-3.1 70B teacher into a Gemma-2 9B student. The goal was to reduce inference latency while maintaining response quality for Indonesian-Javanese customer support. The pipeline includes 8 augmentation strategies that expand the base dataset ~5x, category-aware balancing across ~20 topic types, and multi-layer anti-fabrication validation that rejects hallucinated URLs, invented payment methods, and fabricated policies. Training runs entirely inside Docker — the GPU server requires no Python installation.
MLOps Retraining Pipeline
Built an automated MLOps retraining pipeline via n8n, enabling iterative model improvement with rigorous evaluation. The pipeline has completed 10+ retraining rounds with A/B validation across 154 structured test scenarios — covering product queries, FAQ answering, intent classification, and out-of-scope handling. Each round produces a diagnostic training report with loss curves, inference tests, and automated recommendations, enabling remote collaboration between analyst and server admin.
Demand Forecasting & BI Solutions
Developed end-to-end demand and sales forecasting pipelines using PyTorch and SQL, enabling better inventory planning across 6 darkstores and 2,240+ SKUs. Designed complementary BI solutions including warehouse simulations and a stock balancing system that optimizes goods distribution across the warehouse network — minimizing holding costs and reducing transfer times.