Production-Ready AI & Software Engineer

What's New

A privacy-first code interpreter for data analysis (local-model experiment) - A privacy-first code interpreter for data analysis (local-model experiment) Read more ->

Now Live in Production

Agentic ERP Automation Platform (ongoing) - An agentic automation layer on a live commercial cloud ERP, running the full procure-to-pay chain for a multi-entity manufacturing group: AI vision extraction of vendor invoices, statistical invoice-to-receipt matching, automated purchase orders and purchase invoices written into the ERP, and a three-gate payment flow with enforced separation of duties that executes real bank transfers. Built with no ERP sandbox, so every safety property is engineered: GET-only mirror for all reads, a seven-layer write-guard with idempotency and read-back verification, dry-run by default, canaried one document at a time. 43 read-only AI tools on an in-app assistant. Read the deep dive ->

In Progress

EnergyLM-7B — LLM Fine-Tuning & Alignment — End-to-end LLM fine-tuning pipeline: QLoRA SFT on Qwen2.5-7B, DPO vs ORPO alignment comparison, CoT distillation, 10-benchmark evaluation, AWQ/GGUF quantization. 20K synthetic energy-domain dataset. $0 budget — 100% free compute. View project →

Currently Building

ForceX AI — Indonesia's first AI-for-energy platform. 12 physics-informed AI products for nuclear, geothermal, oil & gas, and renewables. GeoForce v1.1 deployed with R²=0.997 accuracy. Explore the platform →

Overview

I am Robi Dany Riupassa, a software engineer and AI specialist building production-grade AI systems — from physics-informed energy platforms to agentic RAG chatbots and enterprise MLOps pipelines.

My work spans:

  • Agentic Enterprise Automation - Production ERP integration, guarded and idempotent write-back into live financial systems, human-in-the-loop approval flows, AI document extraction, MCP tool servers
  • LLM Fine-Tuning & Alignment — QLoRA SFT, DPO, ORPO, CoT distillation, reward modeling, synthetic data generation, AWQ/GGUF quantization, lm-evaluation-harness
  • Agentic AI & RAG — LangGraph multi-agent orchestration, self-reflective retrieval, hybrid search, multi-provider LLM fallback chains
  • Physics-Informed ML — PINNs, CNN surrogates, GNNs for energy sector simulation (ForceX AI, 12 products)
  • MLOps & Model Deployment — Automated training pipelines, model registry, 24 production models, CI/CD with GitHub Actions
  • GPU Computing — NVIDIA CUDA, PyTorch cu128, Docker GPU passthrough, VRAM budgeting, Blackwell sm_120 compatibility
  • Cloud & Infrastructure — Google Cloud Platform (Cloud Run, Cloud SQL, Secret Manager), Docker, Kubernetes, Kafka, Terraform
  • Computer Vision & Video Analytics - Detection and multi-object tracking (YOLO11, ByteTrack), behavioural event modelling from trajectories, zone geometry design, CPU-only reproducible pipelines
  • Multimodal AI — NLP (spaCy, Transformers), Vision (CLIP, YOLOv8), Audio (Whisper), Vector Search (pgvector)
  • Published Work — Models and datasets on HuggingFace, AI testing framework on npm (@robi-atp/cli)

Now Live — Agentic ERP Automation Platform

Agentic ERP Automation Platform | Ongoing, live in production

An agentic automation platform layered on a live commercial cloud ERP platform, automating the full procure-to-pay chain for a multi-entity food manufacturing and hospitality group. Four connected pillars (purchasing, production, sales, payment), 87 API endpoints, 43 read-only AI tools, 15 containerized services, and 5 guarded ERP write capabilities that create purchase orders, goods receipts, purchase invoices, vendor masters, and payments in a live financial system.

The engineering constraint that shaped everything: no ERP sandbox and no test tenant, with real company money behind every write. The answer was a GET-only nightly mirror serving all reads, a seven-layer write-guard (host guard, mirror refusal, per-capability flag, dry-run shadow mode, deterministic idempotency codes, pre-check with sanity caps, read-back verification), and a rollout that canaried each capability on a single real document before enabling it. The AI proposes; a human commits; the execution layer holds every safeguard.

Stack: FastAPI · Next.js · PostgreSQL · Claude (headless CLI + vision) · MCP · Docker Compose · Alembic · Nginx

View Project →

AI-Native Applications — TransForce & CutForce

TransForce | CutForce | both deployed

Two applications on one architectural bet: a human and an AI agent editing the same live state, through an MCP tool layer rather than code generation, driven by a headless Claude Code agent on a subscription.

TransForce does chat-driven Monte Carlo radiation transport on OpenMC: describe a reactor problem, and the agent builds the CSG geometry, materials, source and tallies, runs it, and reports k-eff with uncertainty. It reproduces published papers end to end and reports the bias in pcm. The agent never writes Python, only a typed ModelSpec that a deterministic driver translates. 27 simulation tools, depletion, weight-window variance reduction, MGXS.

CutForce is a browser video editor where the agent is a second cursor on the timeline: 43 MCP tools, live WebSocket sync, undo that works across both editors, an embedded terminal, and an ffmpeg export engine that compiles keyframes into ffmpeg expressions.

TransForce → | CutForce →

Flagship — ForceX AI

ForceX AI — AI-for-Energy Platform | forcex-ai.com

12 physics-informed AI products (PINNs, CNN, GNN, LSTM, RL) with LangGraph agent orchestration. GeoForce CNN surrogate deployed at R²=0.997. Model and dataset published on HuggingFace. 484 passing tests.

In Progress — EnergyLM-7B: LLM Fine-Tuning & Alignment

EnergyLM-7B — LLM Fine-Tuning & Alignment Pipeline

End-to-end LLM training pipeline: fine-tunes Qwen2.5-7B on 20K synthetic energy-domain instructions using QLoRA SFT, aligns with DPO vs ORPO comparison study, trains a reward model, and benchmarks across 10 evaluation dimensions. Multi-teacher data generation (Gemini + Groq + OpenRouter), MinHash + semantic dedup, LLM-as-judge quality filtering, AWQ/GGUF quantization, vLLM serving. $0 budget — entirely on free-tier compute.

Stack: PyTorch · Transformers · TRL · PEFT · QLoRA · DPO · ORPO · lm-eval-harness · vLLM · Kaggle T4

View Project → | ForceX-AI on HuggingFace →

Latest — Agentic HR Intelligence Platform

Agentic HR Intelligence Platform

Full-stack AI-powered HR platform that turns raw employee and attendance data into workforce intelligence. Gemini 2.5 Flash agentic chat with 5 specialized tools and up to 10-iteration tool calling, 6 interactive analytics dashboards, multi-factor turnover risk scoring (0-100), automated anomaly detection, and AI-driven contract renewal recommendations with PDF report generation.

Stack: Next.js 16 · React 19 · FastAPI · PostgreSQL · Gemini 2.5 Flash · Recharts · Tailwind CSS 4

View Project →

Computer Vision - Retail Video Analytics

Retail Video Analytics: Customer Behaviour from CCTV

Two retail CCTV cameras turned into behavioural metrics: store interest and walk-in conversion, per-shelf engagement events, and staff-to-customer interaction sessions. YOLO11m detection with ByteTrack identity persistence, running entirely on CPU and reproducible from one Docker command.

The work sits above the detector. Perception is cached so behavioural thresholds can be tuned interactively rather than at 80 CPU-minutes per pass; speeds are normalised by body height so one threshold holds across the frame; shelf assignment uses the torso after an audit showed browsing customers' feet never leave the aisle. An apron-colour staff classifier was built, measured, and rejected for a spatial rule that actually separates the classes.

Stack: Python · YOLO11m (Ultralytics) · ByteTrack · PyTorch CPU · OpenCV · pandas · Docker

View Project →

Enterprise Systems (Private Company)

AI Service Migration: CPU to GPU (RTX 5060) — Migrated multimodal AI inference (Whisper, CLIP, YOLO) to NVIDIA RTX 5060 GPU. 8.7x pipeline speedup, Blackwell sm_120 CUDA fix, VRAM budgeting on 8 GB, zero-downtime rolling cutover.

Enterprise Agentic RAG Chatbot — LangGraph multi-agent RAG with self-reflection, hybrid search (dense + BM25 + RRF), multilingual, Langfuse observability. Production-deployed.

Media Platform with OTT Streaming — 8+ NestJS microservices, multimodal AI (NLP/Vision/Audio), Kafka event streaming. OTT deployed on Google Cloud Platform (5 Cloud Run services, Cloud SQL, Compute Engine).

Broadcast Analytics — MLOps Platform — 24 trained models (6 algorithms x 4 targets), LangChain agent with 7 tools, genetic algorithm optimizer, Prometheus/Grafana/Alertmanager monitoring stack.

AI-Powered Help & QnA System — Production RAG with hybrid retrieval, 5-provider LLM fallback chain, cross-encoder reranking, voice assistant. 75%+ cache hit rate, ≤5% hallucination rate.

Independent Projects

GeoForce — CNN Reservoir Surrogate — Standalone production model replacing TOUGH2 simulation. R²=0.997, 57K params. Published on HuggingFace.

Porto Agent — AI content marketing platform. Generates blog posts from git commits using Gemini AI + LangGraph, with human-in-the-loop approval via dashboard and Telegram.

AI Testing Framework (ATP) — Playwright-like testing for AI apps. Semantic similarity, hallucination detection. Published on npm as @robi-atp/cli.

Recommendation Systems Research — YouTube Two-Tower, Netflix Foundation, and Hybrid architectures compared across 108 hyperparameter configurations.

Browse all projects →

Professional Experience

Developer & Researcher — ForceX AI | Current

Building Indonesia's first AI-for-energy platform — 12 physics-informed AI products, GeoForce deployed, models published on HuggingFace.

AI Engineer (Enterprise Automation) — A Multi-Entity Manufacturing & Hospitality Group | Current

Sole engineer on a production agentic automation platform integrated with the group's live cloud ERP: procure-to-pay automation, AI document extraction, guarded ERP write-back, and a multi-gate payment approval flow executing real vendor transfers.

Software Developer & AI Specialist — A Leading ICT Solutions Provider | Current

Enterprise AI systems for media and broadcast — agentic RAG chatbots, MLOps pipelines, multimodal AI, cloud-native deployment on GCP.

Training Instructor — Government Transformation Academy, BPPTIK KOMINFO

Delivered Fundamentals of Data Science training for government professionals.

View full experience →

Education

Doctoral Research in Physics — Bandung Institute of Technology, Indonesia Passive safety systems for Generation IV Molten Salt Reactors, combining CFD simulation with deep learning. ITB Digital Library →

Master of Science (MSc) in Physics — Bandung Institute of Technology, Indonesia

View education details →

Publications

Published research in physics, computational science, and AI. Google Scholar | Scopus

View publications →

Contact