I'm an AI engineer based in Indonesia. My work focuses on ROCm fine-tuning workflows, training-throughput benchmarks for AMD Instinct MI300X, and open-source ML tooling that the broader community can pick up and use without a corporate cluster.
Most fine-tuning recipes online assume you have an NVIDIA box. The handful that target AMD are scattered across blog posts, forum replies, and abandoned forks. There's a real gap between "this paper trained on MI300X" and "here's a clean, reproducible script you can run tonight."
I close that gap. I write recipes — small, focused, run-locally — that take you from a base model to a fine-tuned checkpoint on ROCm 6.x with no detours. LoRA, QLoRA, full SFT, DPO, GRPO. Every recipe ships with the exact requirements.txt, container image, and configuration that produced the run.
I work alone, I move fast, and I'd rather ship a 50-line script that works than a 500-line framework that doesn't. The audience I write for: independent researchers, small teams, students with cloud credits, anyone who wants to actually train something rather than read about training.
Documenting the rough edges in fine-tuning on AMD — bitsandbytes alternatives, FlashAttention-2 ROCm builds, working DeepSpeed configs. Goal: a developer choosing AMD shouldn't lose a weekend to integration friction.
Recipes that fit on a single GPU at FP16/INT4. The 192 GB on MI300X opens up "just throw the whole model in there" workflows that aren't practical on smaller cards.
Every public result ships with the seed, the package versions, the dataset checksum, and the Docker image hash. If someone can't replay it in 2026 they shouldn't trust it.
Side interest: building lightweight evaluation sets for Indonesian instruction following. Open-weight models still underperform on regional language nuance.
Open to GPU compute partnerships, collaborative training runs, and guest-author opportunities for write-ups about ROCm fine-tuning. I respond within a day.