experience
Production ML systems, agentic workflow orchestration, retrieval, knowledge graphs, document intelligence, model evaluation, and cloud ML infrastructure.
Biomedical signal ML and data science education tooling.
Recommendation systems and AI measurement tooling.
Geospatial ML, routing, PostGIS, Redis, and Kubernetes-backed urban mobility systems.
education
M.S. in Computer Science. DARPA Catalyst Award for Caselet, an AI-powered national security workforce tool.
B.Tech. in Electrical and Electronics Engineering.
publications
- Lucky Verma. Solver-Hard Is Not Model-Hard: A Hardness-Controlled Diagnostic for LLM Constraint Reasoning. arXiv, 2026.
- Lucky Verma and Pratik Yadav. When Top-1 Fails: Calibrating LoRA Monitors for Masked Diffusion LMs. arXiv, 2026.
- Lucky Verma. Weight Decay Regimes in Grokking Transformers: Cheap Online Diagnostics. arXiv, 2026.
- Lucky Verma. When Does Removing LayerNorm Help? Activation Bounding as a Regime-Dependent Implicit Regularizer. arXiv, 2026.
selected public software
- solver-hard-is-not-model-hard: hardness-controlled LLM constraint-reasoning diagnostic and aggregate reproduction artifact.
- grokking-diagnostics: transformer grokking diagnostics artifact repository.
- dyt-composition-study: transformer normalization artifact repository.
- top1-fails-dlm-lora-monitors: masked-DLM LoRA monitor-calibration artifacts.
- Document-Classification-using-LayoutLM: public document-classification repository.
- mwb-linux: Linux client for Mouse Without Borders workflows.
technical areas
Python, PyTorch, transformers, LoRA and PEFT evaluation, hardness-controlled LLM evaluation, constraint reasoning, RAG, knowledge graphs, agentic workflow orchestration, OCR, document intelligence, data pipelines, FastAPI, Docker, AWS, Go, Linux tooling, reproducibility automation, and evaluation harnesses.