Robi Dany Riupassa
I build production-grade AI: agentic systems, LLM fine-tuning, and RAG that ship and scale.
New here? Tap the chat bubble in the corner and ask the AI assistant anything about my work.
About
AI/ML engineer with a physics research background and a track record of taking AI from prototype to production. Focus on tested, deployed, measurable systems rather than demos that break in the real world. PhD-level rigor with production discipline.
What I build
- Agentic AI (LangGraph multi-agent, tool calling, self-reflection)
- LLM fine-tuning and alignment (QLoRA SFT, DPO/ORPO, distillation, AWQ/GGUF quantization)
- Production RAG and chatbots (hybrid search, pgvector, caching)
- MLOps and deployment (Docker, GCP Cloud Run, FastAPI/NestJS, GPU optimization)
- Stack: PyTorch, Transformers, TRL/PEFT, vLLM, LangGraph, Gemini, Next.js/React, PostgreSQL+pgvector, Kafka
Selected work
Full-stack platform where a Gemini 2.5 Flash agent uses 5 specialized tools with up to 10-iteration tool calling to answer HR questions and act on data. Turnover-risk scoring (0-100), anomaly detection, 6 interactive dashboards, automated PDF reports. Stack: Next.js, React, FastAPI, PostgreSQL.
End-to-end pipeline: 20K synthetic domain instructions via multi-teacher generation + semantic dedup, QLoRA SFT, a DPO vs ORPO alignment comparison, then AWQ/GGUF quantization. Trained entirely on free-tier compute ($0).
Retrieval-augmented Q&A with hybrid search and aggressive caching: 75%+ cache hit rate and hallucination held to 5% or less through grounding and self-reflection.
12 physics-informed AI products (PINNs, CNN, GNN, LSTM, RL) coordinated by a LangGraph agent. The GeoForce CNN surrogate hit R2=0.997 with only 57K parameters. Published on HuggingFace, 484 passing tests.
Migrated a multimodal pipeline (Whisper, CLIP, YOLO) to NVIDIA GPU with VRAM optimization for an 8.7x end-to-end speedup. Enterprise media/broadcast AI infrastructure.
An open-source testing framework for AI apps (Playwright, but for LLM/agent behavior), published on npm as @robi-atp/cli. Catches regressions in AI outputs before production.
Let's build something that ships.
Available for new projects now. Typical first reply within a few hours.
Start a project