AI Wearables — Smart Glasses Capture & Media Quality (2024–Present)

Meta AI Glasses run on a custom Android stack with real-time capture requirements across 7+ hardware platforms. I joined in 2024 to build the first quantitative quality measurement infrastructure and AI-powered engineering tooling.

• Built two-stage LLM-powered RCA workflow reducing debug MTTR 90% (10 hrs → 1 hr) — comparative prompting with passing runs as few-shot reference isolates anomalous failure patterns; stage 2 generates ranked mitigation paths; human-in-the-loop design preserves engineer judgment; adopted across 3+ teams.

• Identified and owned 0→1 media capture quality workstream — team had no visibility into capture quality before this; defined 5 bad capture criteria (frame drop ratio, frame freeze, audio length, frame rate mismatch, stabilization failure) with DS, DE, PM, and IQ team; designed full Python logging pipeline (iOS & Android); caught 3K video capture frame drop at fleet scale pre-release; expanded from 5→7 device models.

• Diagnosed and resolved 95% Android telemetry drop — traced traffic layer-by-layer to identify naming format mismatch silently dropping legacy devices; unblocked measurement library bugs via sample video reproduction.

• Led Camera2→Camera3 migration for next-gen Meta AI watches (Android 16) as sole engineer — zero-downtime co-deployment across ~40 Kotlin modules, unblocking all future AI camera development on Android 16.

• Built E2E test infrastructure across 21 platforms — cut capture regressions 9→2/month; 52 regression-blocking tests, coverage 79%→100%.

• Owned operational reliability — 10 on-call rotations, SEV remediations including restoring voice-command capture; zero incidents at consumer launch; mentored 3 engineers to promotion.

Ads Monetization — Multi-Destination Lead Generation (2021–2024)

• Stepped into MDLG Tech Lead role and turned around a blocked workstream — introduced revenue-split A/B testing dimension enabling 4 ML workstreams to achieve stat-sig wins previously undetectable; collaborated with ML/DS teams on ad targeting model improvement driving $34.97K/day (2× estimate); grew MDLG $320K→$468K/day serving 100K+ advertisers daily; repaired damaged partner team relationship within 1 week by resetting shared guardrail metrics and experiment ownership.

• Built experiment measurement infrastructure adopted org-wide — introduced revenue-split dimension to A/B testing framework; standardized holdout strategy; designed two-layer holdout preserving long-term measurement integrity while maximizing experiment power across concurrent feature launches.

• Pioneered MDLG L1→L2 entry point migration as lead engineer (React/TypeScript) — first to implement Meta's new store/reducer/component framework at the ad set level; co-defined data model standards reducing future ad type setup from 8 weeks → 2 weeks per engineer; contributed to 1,700% Messenger Destination ad revenue growth ($1K → $18K/day).

• Owned Click-to-Direct Lead Gen (CTDLG) 0→1 as sole product owner (React/TypeScript) — identified gap: businesses couldn't drive direct IG Messaging conversations from ads; executed two-phase roadmap across dominant and second-largest ad creation flows, driving adoption via upselling and defaulting; scaled to $61K/day (2× original estimate).

• Led org-wide code quality as #1 contributor in company-wide engineering quality sprint — 134 fixes (20% of org total), ~1,500 eng hrs saved, $69K/yr compute savings.

Bloomberg's terminal infrastructure processes billions of financial data events daily. I worked on the core telemetry and configuration systems that keep that pipeline reliable at scale.

• Architected C++ real-time telemetry framework processing 2M data points/sec across 750 machines (<1% overhead); automated anomaly detection cut incident impact 80%; zero-downtime deployments at 100% fleet availability; served 10K+ financial terminal clients.

• Built distributed SQL config management system across 750-machine cluster — thread-safe concurrent validation eliminated silent config-drift bugs.

• Course Assistant for "Bayesian Models for Machine Learning"

• Designed anomaly detection algorithms for behavioral information from sensor data of smart light switches in a senior health center and pre-emptively notified caretakers of emergencies.

• Reduced false positive by 56% compared to competing products.

• Published research result in "IEEE International Conference on Big Data Intelligence and Computing," 2016.