Senior AI Product Manager

Co-led Bixby's shift from a reactive command-executor into an autonomous agent that plans and executes multi-step requests. The platform is a multi-agent system: a supervisor agent orchestrates 35 specialized domain agents over a LoRA fine-tuned LLM, planning a sequence of actions and chaining them to complete a complex request end-to-end across apps and the web.

Shipped on the Galaxy S26 (2026), lifting complex-task completion by 15%.

Owned the recovery layer no single team owned: the "unhappy path." For low-confidence intents, a Disambiguation Agent uses contextual cues to predict what the user actually meant and reduce conversational friction. For genuinely out-of-scope queries, a Perplexity-backed Unsupported Agent provides a graceful fallback instead of a dead end.

Together these expanded product coverage by roughly 75% and cut the overall failure rate from 25% to about 6%.

Voice assistants were "digital amnesiacs": great at "set a timer," useless at "what was that restaurant my friend recommended in our chat last week?" I championed a "Second Brain" vision: a personal intelligence layer that sits securely over the user's own data.

The Comprehensive Answer agent uses on-device Retrieval-Augmented Generation: retrieving relevant snippets from Messages, Notes, and Calendar, then synthesizing them with an LLM into a single, human-like answer. I made 100% on-device processing a non-negotiable requirement so personal data never leaves the phone.

It handled 25%+ of queries that previously failed with "No Answer Found," reached 35% monthly adoption within six months, and validated the broader Second Brain roadmap.

Quality was measured on a subjective 1 to 5 "naturalness" scale that every tester read differently, turning weekly quality reviews into opinion battles. I invented a 7-point, binary pass/fail framework (Accuracy, Relevance, Conciseness, Harmlessness, Coherence, Response Length, and View Correspondence) that removed the ambiguity entirely.

I turned it into a weekly dashboard that stack-ranked agents by objective score, drove a department-wide 6.5/7 release bar (shipping gated on quality, not dates), and routed bugs to the right team with clear feedback. A parallel test showed the binary system had ~50% lower variance between testers than the old scale.

Core-domain quality rose from 5.3 to 6.7/7 within a single quarter; the framework became the company-wide release standard, backed by a CI-integrated 5,000+ query benchmark for automated regression tracking.

Replaced a brittle, manually-curated chat system with a modern LLM stack. I split chat into nine intent categories (greetings, emotional, brand protection, device info, emergency, general, and more), each with its own prompt strategy and response style (e.g. emotional replies don't push for follow-ups; greetings are more interactive).

Responses route across three tiers: an in-house LoRA-tuned model for personality, brand, and emotional bonding; a prompt-engineered Perplexity path for brand-advocacy and special cases; and general Perplexity as a fallback for open Q&A and misclassifications.

Shipped in 14 languages on the Galaxy S25 (2025), doubling utterances per user and lifting DAU/MAU by about 50%.

Device Settings had a 40% task-failure rate. The team assumed speech recognition was at fault; I manually analyzed 1,000 failed utterances and found ASR was correct in 80%+ of them. The real problem was semantic: users said "turn on night mode" while the system looked for "Eye Comfort Shield."

I proposed a RAG system that indexes unstructured help content (manuals, tips, support queries) into a vector database and grounds the LLM's answer in the retrieved docs, chosen over fine-tuning for scalability and lower hallucination. A proof-of-concept on a single device manual answered 50 of the top 100 support questions and got the project greenlit.

After launch, Device Settings failures dropped from 40% to under 10%, deflecting 15,000+ Tier-1 support tickets a day.

Partner with Legal on every release and own privacy-by-design review of personal-data handling, including biometric voice data and on-device sensor collection, across the US, EU, and Korea.