Experience
What I've built, and where.
Beyond Horizon Labs
Post-training agents for continual learning
Beyond Horizon Labs is an applied AI lab working on agents that learn from their own execution traces. An agent deployed in a specific domain generates a huge number of these traces, and nearly all of it is discarded. I have a background in consumer ML, where almost every model learns from user behavior, so leaving that signal unused was hard to justify. We are building a recipe for learning from it with on-policy self-distillation, where the agent trains on its own rollouts and each round of deployment produces the data for the next. The goal is an improvement loop that compounds on itself instead of a person patching prompts after every failure.
On the product side, we 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. Every edit and approval it collects is training data for the recipe, so the product and the model improve each other. We raised a pre-seed from Afore Capital to build it.
AIDA
Engineering long-horizon, domain-specific agents
I led AI at AIDA and built a proactive agent for enterprise sales. The work was long-horizon by nature: tasks unfolded across email, meetings, calendars, CRM records, and documents, state was incomplete or contradictory, work persisted across sessions, and every external action needed an approval boundary. The agent resolved that context continuously into commitments, risks, and executable follow-up work, then prepared the work before anyone asked for it. I ran product discovery directly with customers, learning where the agent earned trust and where it wasted their time, and the changes that came out of those conversations lifted engagement significantly and helped take the product from early pilots to six-figure ARR.
I designed the agent harness myself, with retrieval, structured extraction, tool calling, persistent memory, asynchronous task pre-execution, and approval-aware execution. Rather than routing everything through a single general-purpose loop, I built workflow-specific task agents for email, calendar, meetings, documents, and CRM operations.
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.