Experience
What I've built, and where.
Beyond Horizon Labs
Post-training agents for continual learning
I founded Beyond Horizon Labs to explore how agents can learn from execution traces instead of requiring people to patch prompts and workflows after every failure. I built Artifact, a proactive agent for Chiefs of Staff that inferred intent from meetings, email, and calendars, then prepared and executed multi-step work behind a human-approval boundary. Its architecture combined continuous context ingestion, structured extraction, retrieval, tool use, asynchronous execution, memory, and outcome-based evaluation.
My current research turns corrections, approvals, rejections, and task outcomes into trajectory-level learning signals for multi-turn, tool-using agents. I am comparing on-policy self-distillation with GRPO-style reinforcement learning for continual adaptation, personalization, and distribution shift.
AIDA
Engineering long-horizon, domain-specific agents
In 2024, building a dependable long-horizon agent was primarily a systems problem, not a prompting problem. Sales workflows unfolded across emails, meetings, calendars, CRM records, and documents; state was incomplete or contradictory, tasks persisted across sessions, and external actions required careful approval boundaries. I led AIDA’s AI effort and built a proactive agent that continuously resolved this context into commitments, risks, and executable follow-up work.
I designed a custom agent harness with retrieval, structured extraction, tool calling, persistent memory, asynchronous task pre-execution, and approval-aware execution. Rather than relying on a single general-purpose loop, I built workflow-specific task agents for email, calendar, meetings, documents, and CRM operations, along with regression evaluations based on simulated users and rubric-driven LLM judges.
Apple
SFT, semantic retrieval, and curriculum learning
I led the zero-to-one prefix-search system for the global launch of Apple Music Classical. Classical search is an unusually difficult semantic-retrieval problem, shaped by long-tail entities, aliases, multilingual metadata, spelling variation, and distinctions between composers, works, conductors, recordings, and performances. I designed a hybrid architecture combining a deterministic catalog index with a domain-adapted T5 1.1B model trained through supervised fine-tuning.
I trained dual-encoder representations for semantic retrieval and personalized ranking using co-play, co-click, content, user, and query-document supervision. The work focused on combining heterogeneous behavioral and content signals into a shared embedding space, with representation quality evaluated using nDCG and MRR across a catalog of more than 100 million songs.
I also developed curriculum-learning methods for multilingual structured prediction, ordering synthetic, weakly supervised, and human-labeled examples by source quality. Training data combined back-translation, user-engagement signals, and search-backend signals, reducing parse error rate by an average of 2.5% across more than 10 locales.