I’m a Research Engineer at MBZUAI, working on ML systems. My main focus is cross-vendor mismatches between AMD and NVIDIA GPU platforms.

I started with language model architectures and training objectives, keeping their systems implications in mind. Over time, I found myself more drawn to how we organize computation, use memory, and make training more efficient. That interest now guides my research, and I’m looking for PhD opportunities in ML systems.

I’m interested in the interaction between training algorithms and the systems that run them: how we distribute work across devices, use available compute, and understand differences between GPU platforms.

My coauthored work includes COPUS, which studies adapting batch size and parallelism together during language model training. It connects to a question I want to keep exploring: how can model training make better use of the hardware available to it?

My language model research includes Token Order Prediction, published at ICML 2026. In this work, we use the relative order of upcoming tokens as an auxiliary training objective. I also coauthored Softpick, published in Findings of ACL 2026, which revisits attention normalization to address attention sinks and massive activations.

These projects form the language modeling background I bring to ML systems research. I’m interested in how choices in architectures and training objectives interact with their implementations.

I contribute to Unsloth, with merged work across its core library, Unsloth Zoo, and training notebooks. My contributions address the practical details behind model training: kernel data types, compiler settings, library compatibility, and saving trained models.

In the core library, I’ve contributed fixes for data-type mismatches in Triton cross-entropy kernels, compatibility with changes to TRL’s trainer API, and saving vision-language models in 4-bit. In Unsloth Zoo, I also contributed a compiler configuration fix for a vision-language model compilation failure.

I also contribute training examples, including Qwen3-VL fine-tuning and GRPO notebooks. This work connects my language modeling background with the implementation and systems questions I want to explore more deeply.

I also keep hands-on learning projects: triton_inline for GPU kernel programming with Triton, and learning_parallel for exploring tensor and data parallelism. They’re places to work through the implementation details behind model training as I deepen my systems knowledge.

I coauthored COPAL-ID, a NAACL 2024 benchmark for commonsense reasoning grounded in Indonesian culture and language. It evaluates both standard and colloquial Indonesian, with attention to local knowledge and cultural context.

I’m starting a research blog as Erland Hilman Fuadi (Edd), where I’ll share notes on ML systems, efficient training, and what I learn along the way. First posts coming soon.

You can find my publications and background on my website, or get in touch about research and PhD opportunities.