AI Product Manager

The vision was to run Bixby entirely on-device for speed, privacy, and offline use, but a lift-and-shift of 3,000+ server-side goals was impossible on a phone. As founding PM, I scoped an impossible vision into a shippable V1 through three judgment calls:

(1) NLU: usage data showed 20 core goals (alarms, flashlight, and the like) covered 50% of all Bixby usage, so I de-scoped the model to just those, sacrificing the long tail to make it feasible. (2) ASR: non-negotiable quality, allocating the size budget for a 200MB model at 99% accuracy. (3) TTS: pre-load only the single default brand voice that most users actually used.

Shipped on the Galaxy Z Flip (2021): 90%+ adoption in the US and Korea, ~10% of global Bixby traffic handled on-device, 35% server-cost savings, and ~30% lower latency. Won "Best Feature in AI Team" (2021).

An engineer on my TTS team prototyped voice cloning from ~18 recorded sentences. The fast path was server-side processing, but voice is intensely personal biometric data, and privacy ("always listening") was already Bixby's top user complaint.

I advocated against the faster server approach, arguing the long-term cost of lost trust outweighed being first to market, and won a ~6-month delay to build the entire pipeline 100% on-device: no voice data ever leaves the phone, zero logging, privacy by default. We launched iteratively via Bixby Labs, leading with the safest, most compelling use cases: Call Screening and Bixby personalization.

Reached 20M users shortly after launch, earned positive privacy-focused press, and set the on-device gold standard for sensitive-data AI features at Samsung. (Global media coverage.)

I conceived an on-device ASR feature to answer and transcribe calls for spam screening, but it depended on the Phone App team, whose goal (serve all 1B Samsung users) conflicted with mine (drive Bixby engagement via Bixby branding and registration).

Rather than force my goal, I created a shared one: a best-in-class, privacy-first screening feature for all Samsung users, with Bixby's AI as the foundational tech. I made real compromises: dropped "Bixby" from the name (just "Text Call"), de-scoped to 100% on-device ASR with a simple preset-message UI (no Bixby registration needed), and gave the Phone App team equal credit. I embedded my lead engineer with their team to co-design a secure real-time audio path.

Launched on time with 50M+ uses and an NPS lift of ~20 points. This was the first time Bixby's AI powered the whole Samsung ecosystem, and a foundation for what became Galaxy AI.

Custom Wake-up (name Bixby anything) was exciting but imperfect: an 8% false-reject rate versus 2% for "Hi Bixby," and only two languages. A full launch risked the brand; killing it wasted the innovation.

I proposed a third path: Bixby Labs, an opt-in beta platform with clear expectations, a server-side kill-switch, and built-in data collection, letting us ship quickly but safely and learn from enthusiastic early adopters. I had to win a backbone debate with the Head of QA, who opposed shipping anything below the quality bar; I argued that for novel AI, a controlled real-user release is the only way to reach the bar.

Custom Wake-up hit 17% activation and 25% higher weekly actives vs. control, with positive press, and Bixby Labs is now the org's standard launch-and-learn channel for all experimental AI features.

The top user request was to unlock the phone by voice, but literal voice-unlock was impossible, since Android biometric policy requires 99.9% accuracy that voice can't meet. I reframed the real need: users didn't want to unlock, they wanted to get things done hands-free.

I categorized 100+ Bixby actions into three risk tiers: low (toggle flashlight), medium (call mom, read last message), and high (delete a contact, turn off mobile data). For each, I designed safeguards: masking the caller number on the lock screen, a clear opt-in "use Bixby while locked" setting, and voice-match as step-up authentication for high-risk actions only.

Result: 40% of all Bixby commands now execute over the lock screen. I also added Continuous Conversation (on-device audio rejection, speaker verification, and a contextual turn classifier) for wake-word-free follow-ups.