Data &
ML Engineer

A research project in collaboration with Workhuman. Extracts promotion patterns from employee award messages using an LLM pipeline, validated with statistics.

• Extracted promotion-related signals from employee award messages using an LLM pipeline (Workhuman collaboration)
• Validated reliability with Bootstrap CI, effect size, and 4-model cross-validation

Customized BLIP-2 architecture by replacing OPT with Phi-1.5 to meet a sub-3B parameter budget. Fine-tuned Q-Former with multi-stage training. Ranked 43rd / 242 teams (top 18%).

• Customized BLIP-2 with Phi-1.5 LLM under 3B parameter and pre-2024 model constraints
• Multi-stage fine-tuning (captioning → VQA → reasoning → closed-ended); 0.7665 weighted accuracy, top 18%

Multi-class image classification of used car images into 396 make/model/year classes. Background removal with YOLO, ConvNeXt-Base fine-tuning, and TTA + soft-voting ensemble. Ranked 134th / 748 teams (top 17.9%).

• Built fine-grained image classifier for 396 car make/model/year classes (top 17.9%, Hecto AI Challenge)
• Background removal with YOLO + ConvNeXt-Base fine-tuning + TTA + soft-voting ensemble

Designed and implemented a relational database for a pet hotel business using MySQL. Covers ER modeling, schema normalization, and complex SQL queries (JOIN, subquery, EXISTS).

• Designed and implemented a relational database for a pet hotel using MySQL
• ER modeling, schema design, and complex SQL queries (JOIN, subquery, EXISTS)