I have extensive experience managing the end-to-end AI training lifecycle, which includes both large-scale compute orche
I have extensive experience managing the end-to-end AI training lifecycle, which includes both large-scale compute orchestration and hands-on, manual data annotation. As a Lead Researcher, I have overseen training runs for the Nemotron-Terminal-14B model utilizing ROCm 7.0 and successfully trained five separate Mixture-of-Experts (MoE) models based on Mistral architectures, supported by a local high-density GPU laboratory. Regarding data curation, I have extensive experience in both manual data labeling and automated dataset filtration. I frequently conduct hands-on, manual annotation to establish high-fidelity, gold-standard seed datasets (specifically for complex code-generation and reasoning tasks) where automated labeling is insufficient. My manual labeling experience includes: Ground Truth Curation: Hand-labeling and sanitizing initial seed datasets of several hundred high-quality prompt-response pairs to serve as the baseline for fine-tuning. RLHF Preference Labeling: Manually evaluating, grading, and labeling model outputs to train reward models and align model behaviors with specific human-preference guidelines. Annotation Taxonomy & Audits: Authoring the physical labeling guidelines and edge-case taxonomies, and conducting manual quality-assurance audits on automated or crowd-sourced labeled outputs to measure inter-annotator agreement and resolve label noise [1]. This combination of hands-on, manual labeling precision and systems-level algorithmic filtration ensures that the training datasets meet the exact quality standards required for production-level LLM performance.