AI Engineer (LLM Fine-tuning & Data Labeling)
Led a custom few-shot learning pipeline utilizing LLaMa 3 (8B) with 600 fine-tuned examples for the extraction of job roles and skills from unstructured job advertisements. Achieved 95% accuracy and eliminated the need for two months of manual preprocessing, supporting a national labour intelligence AI system. Work included strategy design for lexical and semantic matching and integrating inference-based techniques to process 167k job records efficiently. • Collected and curated labeled datasets using random 8-sample selection per session. • Developed labeling instructions and evaluated data quality for model tuning. • Automated semantic chunking and hybrid search methods as part of data preparation. • Collaborated closely with government stakeholders for requirement refinement and real-world deployment.