Experience

Architected Translation First framework using PyTorch with agent-based approach separating perception and reasoning, outperforming 32B models on MMMU and MIA-Bench benchmarks (SeeingEye, arXiv 2025). Designed reinforcement learning training framework for temporal reasoning tasks with three-stage curriculum and dynamic reward systems (Time-R1, arXiv 2025). Built production-grade evaluation infrastructure to benchmark multiple LLMs on knowledge-intensive VQA datasets.

Developed an internal prototype for automated short-video generation using Python, FastAPI, Node.js, and Docker, integrating topic-based script generation with media processing and video assembly end to end. Adapted open-source workflow-automation and AI video-generation components into a containerized demo, exposing backend services through REST APIs and validating the pipeline with lightweight models.

Engineered end-to-end ML pipeline using Python, scikit-learn, and OpenCV, processing 1000+ high-resolution images with automated data ingestion and model training, achieving 20% accuracy improvement through iterative optimization.

Built a scalable LLM evaluation pipeline with Python and PyTorch over 3,000+ examples, with automated data validation and batch scheduling. Designed model scoring and monitoring modules (REST API and logging), improving scoring efficiency by 20% through iterative optimization and shortening the training cycle by 15% with better version management.