Hi there 👋

I build end-to-end AI systems, from RL research and agentic pipelines and production deployment.

Agentic Systems

Financial Agentic RAG — LangGraph agent routing financial report queries across Qdrant vector search and SQLite, with CRAG-style relevance loops and SSE streaming to a React frontend.

VoiceNav — LLM browser automation agent with Text vs Vision planner ablation (+17pp task success), two-layer failure attribution, and CDP-based live browser streaming.

Research

Web World Model — RL training system for web navigation agents. SFT + GRPO on Qwen2.5-3B on 2×A100 with vLLM rollout serving, 150K synthetic samples — +44% over CoT, 4.4× faster than Tree Search.

Spatial VLM Investigator — Spatial reasoning in VLMs via CoT and RL fine-tuning. GRPO beats SFT on OOD generalization (3.17% vs 12.03% ID-OOD gap).

Academic Knowledge Graph — End-to-end KG pipeline: crawling → ontology construction → BERT+BiLSTM+CRF NER → Neo4j → semantic retrieval.

Knowledge & Infrastructure

ML Inference on Kubernetes — PyTorch training Job + inference Deployment on GKE, shared PersistentVolume, liveness/readiness probes, LoadBalancer REST API.