About Naol Haase
Naol David Haase is a Machine Learning & AI Systems Engineer based in Germany. He designs and pre-trains novel architectures (recursive models, state-space backbones), fine-tunes models with SFT and QAT recovery, and engineers production-grade MLOps pipelines and autonomous agent runtimes on NVIDIA Grace Blackwell, DGX, and AMD ROCm hardware.
Work & Education Experience
- 2024: GitHub Student Developer — Entry into Systems & AI Stack
- 2025: Google Hackathon — 1st Place Winner, AI & Agents
- May 2026: 4PL Intermodal GmbH — AI & MLOps Engineer (IMSLOT)
- 2026: ML Systems & SFT Pipeline — Fine-Tuning, QAT & DGX Spark
- 2026: Machine Learning Systems Engineer — Architecture Research & Production MLOps
Key Projects & Research
Projekt IMSLOT (4PL Intermodal GmbH)
May 2026Production document-extraction & MLOps pipeline for 4PL Intermodal GmbH. Converts unstructured freight documents into verified JSON using Qwen3 fine-tuned with QLoRA & recovered via NVFP4+QAT, served via vLLM with prefix caching on NVIDIA DGX Spark.
Recursive SSM Architecture (Theimpossible)
Sep 2026Pretrained an 84M-param dual-timescale recursive model with a Mamba-2 state-space backbone from scratch on 1.4B+ tokens (FineWeb-Edu). Built with Triton SSD kernels on AMD ROCm (RX 9070 XT) and distributed vast.ai nodes with mechanistic ablation analysis.
Whole-Brain Emulation (FlyWire)
Jul 2026Simulated the real FlyWire connectome (138k biological neurons, 5M synapses) driving a 42-DOF physical fly body in MuJoCo. Built custom event-driven sparse kernels on an AMD RX 9070 XT (ROCm) with real descending motor neurons.
Autonomer Trading-Agent
Apr 2026An algorithmic trading agent utilizing Gemini 2.5 Flash & the Alpaca Trading API. Operates completely autonomously on a $100,000 paper trading account with self-reflection loops, yielding a verified profit of €500+.
Angewandte Robotik (Thor & Moveo)
2025 - 2026Designed, 3D-printed, and programmed functional robotic arms based on the Thor and Moveo models. Features precision kinematics control using NEMA-17 stepper motors and custom microcomputer systems.
Google Hackathon Winner
2025Won 1st place in a Google Hackathon by building a novel AI agent prototype that automates system workflows and integrates multimodal LLMs for real-time task reasoning and self-reflection loops.
