I am Sungin Kim.

Since 2018 I have worked as an employed firmware engineer, on FPGA and embedded systems. The silicon I have worked on includes Microsemi PolarFire, Stratix 10 through Quartus, Xilinx ZCU106, and other Zynq SoC families, on development boards and on customer boards. Before I built this lab I was doing that work. This site is how I show the AI systems I build now, so a prospect can see who they would be hiring: a GPU homelab, trading-signal research, and the experiments I killed. There is no client list here. The integrations index is empty because there are no paid engagements to file.

I started the box because cloud rate limits kept interrupting inference tests on a 30B-parameter model. What exists now is four RTX 5060 Ti 16GB cards (64GB of VRAM), a Ryzen 9 9950X, and 128GB of DDR5, on an open mining frame. It cost $5,638.79. Almost none of that was the original plan. That write-up is on this site.

The host runs Proxmox. All four GPUs go to one research container - 24 CPUs, 80GB of RAM - which trains, backtests, and serves inference. I did not give each card its own container. The PCIe lane map decides which pairings are allowed. That is a separate post.

I specced the machine for language-model inference. After seven months of daily use it was doing a different job: three GPUs training TFT and XGBoost models for a crypto trading-signal system, and the fourth serving inference for backtests, paper trading, and live runs. One published trading experiment was an LLM forecaster that looked well calibrated and still lost to the market in every bucket that mattered. I pre-registered a kill gate and killed it. Another write-up is a look-ahead bug that made a backtest look worse than it was. No live money was on that killed forecaster.

When the hardware misbehaves, that becomes a post too. The rest of the record is the experiment log.