SINCE 2026
Meta
Software Engineer (Production Engineering)
- The messy middle
- Benchmarks & honest numbers
- Research to production
- The founder side of building
Production Engineer · Meta · Bucknell University
Chang Min Bark
The engineer who lives in the messy middle and measures every win.
Chang Min Bark is a production engineer on Meta's Database Platforms team, keeping high-availability MySQL ClusterSets online for enterprise customers. He came up through startup backends at Hangry, where he cut notification latency 30ms → 1ms, and open-source infrastructure — a merged SGLang scheduler, Kubernetes inference-perf benchmarks, a Blockly plugin. Before any of it he poured house foundations with Habitat for Humanity in Indonesia and turned waste cooking oil into donated soap.
- 99.99%Meta database-platform SLA (125+ customers)
- 176%SGLang decode throughput, 4 requests
- 500+Meta HA databases, mass-failover automated
- 97%Hangry notification latency cut (30ms→1ms)
HAPPENING NOW
STANDING TODAY
HOW HE THINKS5 BELIEFS
Five working beliefs, each with a quote or a merged PR behind it.
THE IMPACT
Pull one out — each carries the number that proves it worked.
EVERYTHING ELSE · 5 WORKS
THE TIMELINE
From pouring house foundations in Surabaya to scheduling the CPU–GPU handoff inside a 32k-star LLM engine.
The whole record, newest first – tap any moment and Aster tells the story.
2026–PRESENT
The messy middle: LLM inference at scale
Now full-time on Meta's Database Platforms team, Chang Min works the handoff where a request crosses machines. Outside work he shipped CPU–GPU decode-overlap scheduling into SGLang — a +1,051/−149-line merged PR that lifted throughput ~176% at four concurrent requests — benchmarks Kubernetes inference serving across 300+ LoRA adapters, and adds test-time-training adapters to a frozen Gemma 3.
“The main code logic was written by me. Claude Code was used to polish the code and write comments/docstrings in the code (as well as parts of the PR description). Everything was verified and understood by me.”
2022–2025
Bucknell: research, RAG agents, and shipping end-to-end
At Bucknell (summa cum laude, 3.97 GPA) Chang Min built in every direction: Chiron, a multi-agent RAG teaching assistant that reached the 2025 BizPitch finals; a Blockly multiselect plugin modernized for Google Summer of Code 2024; and a Go notification service at Hangry cut 30ms → 1ms for 20,000+ users. Research with Prof. Thiago Serra seeded the ML half of his work.
“We developed a virtual teaching assistant that allowed students to get real-time assistance and talk with their course material. Once a student asks a question to the TA, the TA will respond and cite its sources, pulling up the material in a PDF viewer in a matter of seconds.”
2016–2021
Surabaya: service, construction, and curiosity
Before software there was concrete. Growing up in Indonesia, Chang Min poured house foundations with Habitat for Humanity (2018–2020), built a village school library near Batam, planted coral-reef frames off Banyuwangi, and turned used restaurant cooking oil into donated soap bars — teaching other teens to do the same. The habit that carries: build something end-to-end, and finish it.