I build applied machine learning systems and public research artifacts around transformer behavior, model monitoring, document intelligence, retrieval, software-development tooling, and knowledge-graph-backed workflows.
My public work focuses on scoped empirical ML studies and reproducible artifacts: controlled transformer normalization experiments, monitor calibration for LoRA fine-tuning in masked diffusion language models, developer tooling, document intelligence, OCR, and local-first automation.
At Eccalon LLC, I work on production ML systems for traceable operational workflows, including retrieval, agentic orchestration, knowledge graphs, and document-understanding pipelines.
recent news
- June 2026: Released When Top-1 Fails: Calibrating LoRA Monitors for Masked Diffusion LMs, with Pratik Yadav, along with public aggregate artifacts and reference scripts.
- May 2026: Released Weight Decay Regimes in Grokking Transformers, with public diagnostics code, aggregate artifacts, Lean checks, and a Hugging Face dataset.
- April 2026: Released When Does Removing LayerNorm Help?, a solo-author empirical study with code, validation scripts, result manifests, and a Hugging Face artifact dataset.
- 2026: Public GitHub portfolio reached 97 repositories, including 50 non-fork repositories and 126 public stars as of June 26, 2026.
- UMBC: Completed an M.S. in Computer Science and received the DARPA Catalyst Award for Caselet, a national security workforce tool.
research interests
Transformer training behavior, PEFT monitoring, masked diffusion language models, developer tooling, evaluation design, retrieval systems, knowledge graphs, document intelligence, and source-grounded ML workflows for regulated operational settings.
The best way to reach me is via email at luckie.verma30 at gmail dot com.