Professional work / 2025
Language Model Training System
Built an end-to-end engineering workflow for a roughly 2B-parameter language model on an 8×H100 cluster, from data preparation to evaluation and inference.
Context
The project tested whether a medium-scale language model could serve a specific task. The core work was not a single training run; it was creating a repeatable chain across data, training, monitoring, evaluation and inference.
My role
- Owned the technical direction and end-to-end training workflow.
- Built preprocessing, tokenization and training data pipelines in PyTorch.
- Organized distributed training and multi-week monitoring on an 8×H100 environment.
- Established evaluation and inference checks for quality and stability.
Engineering judgment
The training scale was roughly 2B parameters. At this stage, traceable data versions, configurations, monitoring signals and evaluation evidence mattered more than pursuing parameter count alone.
Disclosure
This page includes only responsibilities and methods that can be discussed publicly. It excludes the use case, source data, weights, internal metrics and client context.