AI · MLOps · Solutions Architecture
Adam
Oentoro
Machine learning and AI, focused on evaluating whether these systems work and on building, deploying, and running them in production.
01About
From Research
to Production
I studied AI at VU Amsterdam and taught machine learning there for three years. These days I focus on the engineering around AI: evaluating whether these systems actually work, and building, deploying, and operating them so they run reliably in production.
My MSc thesis was as much an evaluation project as a systems one: a deterministic, LLM-as-judge pipeline that scored 270 simulated conversations across eight dimensions and three model providers, measuring how ontology grounding changed their behaviour.
What I enjoy most is the systems side: cloud infrastructure, deployment pipelines, and monitoring that take AI from a notebook to something that runs reliably. I've shipped async ML inference services with FastAPI, Celery and Docker, and a home stack that cut full redeploys from hours to minutes.
Places I've Called Home
02Projects
Selected Work
GPU / HPC
CUDA GEMM Optimization
A GPU matrix-multiply optimization ladder in CUDA, from naive to tiled to cuBLAS, profiled end to end with Nsight.
Naive → tiled → cuBLAS, Nsight-profiled
MLOps
F1 Inference Service
Async ML inference service (FastAPI, Celery, Redis, Docker) predicting Formula 1 finishing positions, with cold-start optimization and leakage-aware feature engineering.
Async inference with cold-start optimization
Evaluation · MSc Thesis
LLM-as-Judge Evaluation Harness
A deterministic LLM-as-judge pipeline scoring 270 simulated conversations across eight dimensions and three providers (GPT-4 Mini, Claude 3.5, Grok 3), measuring how ontology grounding changed model behaviour.
13.9% performance variation discovered between providers
More Work
Evaluation · In Progress
Neural Dynamic
An AI-native assessment platform for technical decision-making roles. Candidates work through realistic scenarios with an LLM in the loop, scored on judgment rather than trivia, with decisions modelled as traceable graph objects.
Reproducible, explainable scoring by design
MLOps / Fintech
Credit Default Scoring
Async probability-of-default scoring service built with FastAPI, Celery, Redis, and Docker, with sync, async, and batch-portfolio endpoints. Logistic-regression scorecard, tests, and CI.
Sync, async & batch scoring, with CI
Research Project
Reinforcement Learning for Robotics
Trained DQL, PPO, and DDPG agents for mobile navigation and manipulation. Successfully transferred policies from simulation to physical hardware.
Sim-to-real transfer achieved
BSc Thesis
Self-Adaptive Systems Profiling
Extended MockSAS with DingNet IoT scenario profile to study multi-armed bandit policies in self-adaptive systems.
Novel IoT scenario for policy benchmarking
Infrastructure
Smart Building Automation
Integrated 170+ KNX devices with Home Assistant. Ansible, Docker Compose, and Tailscale for deployment and secure access.
11 climate zones managed
03Background
Work, Study & Credentials
Work2026–Present
Infrastructure and Home Automation Engineer
Independent
Designed and operated a smart-building platform integrating 170+ KNX devices with Home Assistant.
Work2026–Present
AI & Digital Solutions Consultant
Aktiv Asia Management
Evaluated AI and financial software solutions, produced banking and fintech market analysis for Indonesia & Southeast Asia.
Degree
MSc Artificial Intelligence
Vrije Universiteit Amsterdam
Thesis: a deterministic LLM-as-judge pipeline scoring 270 simulated conversations across eight dimensions and three model providers — measuring how ontology grounding changed their behaviour. Found 13.9% performance variation between providers.
Read the full breakdownWork2022–2025
Teaching Assistant
Vrije Universiteit Amsterdam
Delivered lab sessions and tutoring across Machine Learning, Deep RL, and Information Management courses.
Degree
BSc Artificial Intelligence
Vrije Universiteit Amsterdam
Thesis: exploring reinforcement learning for profiling self-adaptive systems. Extended MockSAS with a city-scale IoT (DingNet) profile.
Read the full breakdownDegree
Started BSc Artificial Intelligence
Vrije Universiteit Amsterdam
Technical Stack
Evaluation
- LLM-as-judge pipelines
- Statistical evaluation
- Benchmarking
Languages
- Python
- Rust
- SQL
- R
- Java
ML & GPU
- PyTorch
- TensorFlow
- CUDA
- Nvidia Nsight
- Stable Baselines 3
MLOps & Serving
- FastAPI
- Celery
- Redis
- Docker
- GitHub Actions
Cloud & Infra
- AWS
- Terraform / OpenTofu
- Ansible
- Tailscale
- Git
04Writing
Blog & Thoughts
View All Writing
12 May 2026
Designing an Ontology for Formula 1
Designing an Ontology for Formula 1: why LLMs struggle with deep, structured questions, and how a domain‑specific ontology can model teams, drivers, cars, results, and evolving regulations to answer nuanced, historically grounded questions that general‑purpose tools cannot
28 Mar 2026
Sustainable AI: Doing AI Without Blowing Up ESG Targets
AI is moving into production just as its energy and water footprint explodes. What is “green‑in” and “green‑by” AI, and how do we get real value from AI without blowing up climate and water targets.