Spearheaded infra unification work using Hack (Meta’s in-house PHP dialect) to eliminate double decisioning in the funnel, reducing up to 42% duplicate requests and 50%+ system complexity, avoiding several potential production incidents, and unblocking LLM adoption in the funnel. Held design reviews and knowledge sharings to keep the team informed.

Led the Lightweight Decisioning work group (4 devs) aimed at lowering the resource consumption of the Holistic Decisioning system, exploiting multiple levers including pre-filtering, Memcached caching, cost-labeling using latency data from Hive and LLM holistics, short-circuiting logic, fan-out limitations, and a dedicated low-latency mode. Resulted in 15% compute cost reduction, 3× more log retention with storage cap unchanged, 80% lower load on model backends, and 60% fewer timeout errors.

Fixed a long-standing “overexposure” problem in the team’s experimentation framework integration, anticipating to support 100+ A/B experiments of policies in 2026. Fixed two debugging tools to save ~2 engineer-week in 2025.

Revamped team onboarding docs with unique perspective, gaining 100% positive feedback from new members and becoming the foundation of the team’s custom AI agent (10% higher AI adoption than sister teams).

Led the development of Bilibili’s content regulation system “Avalon”. Built with Go and gRPC, Avalon superseded ad-hoc filters in 20+ products. It supported 5,000+ peak write QPS using TiDB, Redis, and Kafka.

Integrated 20+ upstream services, such as AI sentiment analysis, text similarity and user relations, using service orchestration. By leveraging NLP models and user interaction data, kept the proportion of content needing human intervention below 10%.

Communicated with the content moderation team and collected requirements there. By developing intervention tools for moderators, increased the efficiency of compliance work by 300%.

Initiated and implemented strategies to maintain a positive community atmosphere. These reduced spamming in Bilibili IM by >90% and the number of creators reporting harassment by >80%.

Developed a strategy framework to enable code reuse and accelerate new strategy rollouts (<1 man-day per rule in 2022).

Migrated services including “Like and dislike” to a multi-site active-active architecture.

Developed the Content Management System (CMS) in a Scrum team using Go, the Gin framework, Amazon S3, and MongoDB. The CMS supported authoring and organizing multimedia courses for 15 product lines and one million paid users.

Developed “CourseScript,” an in-house Domain-Specific Language (DSL) designed for course contents. The DSL parser and encoder supported 100+ activity types and increased content authors’ efficiency by 50%.

Worked closely with the algorithm team to deliver a streaming spoken English grading system, using WebSocket and gRPC. Compared to bulk processing, reduced grading latency from ~10s to <1s.

Designed and implemented company-wide SMS and WeChat push services with availability of 99.99%. Collected and shipped requirements from sales and marketing teams, such as batch promotions and tracing.

Conducted scrum meetings, led code reviews, and mentored junior developers.

Built CooTek’s Voice-over-IP (VoIP) infrastructure by modifying open-source PBX software Asterisk and SIP stack Resiprocate, in C and C++. Added monitoring using OpenTSDB and load balancing using DNS and consistent hashing.

Integrated the VoIP system with CooTek’s user account and billing services using C++ and Protocol Buffers.

Authored “dial plan” scripts in Lua to connect the VoIP system with external users via Public Switched Telephone Network (PSTN). This was TouchPal Dialer’s main feature, which is similar to Google Voice.

Laid the foundation of CooTek’s walkie-talkie app “BiBi” by extending and adapting the PJSIP stack to mobile platforms.

By 2016, CooTek supported 10 million VoIP calls per day, and TouchPal Dialer had more than 100 million MAU.