Experience

Developed AutoForge, a framework that formulates automatic agentic system generation and optimization as a searching problem using Monte Carlo Tree Search (MCTS) to explore agent architectures. The framework incorporates cost-aware optimization by intelligently assigning heterogeneous LLMs based on task-specific capabilities, maintaining strong performance while significantly reducing inference costs.

Selected as a Google Summer of Code 2025 contributor to enhance ChromiumOS's farfetchd service with kernel tracing and replay capabilities.

Designed and implemented an internal front-end component library for Microsoft Edge team.

Led the group's first on-device learning project.

End-to-end incubated FreeInference, an OpenAI- and Anthropic-compatible inference service for open-source, research, and education use. Proposed two LLM routing algorithms, Nimbus and RouteWise, for latency- and cost-aware routing across local GPU deployments, serverless APIs, and multi-provider endpoints.

Co-leading OpenManus-RL, an agentic reinforcement learning framework for fine-tuning LLM agents on environments such as ALFWorld, WebShop, and GAIA. Co-developed AgentDebug, a framework that analyzes LLM agent trajectories with a modular error taxonomy to localize root-cause failures.

Developed BEDEO, a multi-agent ontology-based LLM recommendation framework that enhanced interpretability and recommendation quality for Canada Ontario residents. Integrated ontological knowledge structures with LLM reasoning to improve recommendation transparency.

Led the creation of OasisSimp, a multilingual text simplification dataset for English, Sinhala, Tamil, and Thai. Implemented and fine-tuned mT5 using the MUSS framework for unsupervised text simplification.