Location: San Francisco, CA

Remote: No. SF onsite or hybrid.

Willing to relocate: No

Technologies: TypeScript, Node.js, Python, Rust, distributed systems, event-driven architecture, microservices, real-time systems, multi-tenant, WebSockets/streaming, NATS, AWS (EKS, ECS), GCP/Firebase, Kubernetes, Docker, Pulumi, PostgreSQL, Firestore, MongoDB, Redis, Elasticsearch, Prometheus, Grafana, agent orchestration, LLM integration, RAG, MCP, human-in-the-loop, LLM-as-judge, evals and observability, OpenTelemetry, fine-tuning (LoRA)

Résumé/CV: https://drive.google.com/file/d/13tDwa9HFDxqJVPsYg-9IhlwJ65k...

Email: skinner.cheng@gmail.com

Ava, a real-time captioning platform for one-on-one and group conversations, online and in person. 11+ years, zero to product. I co-founded it and was CTO, but the backend was mine throughout, first building it and later as Backend Lead. What that covered: the live captioning pipeline, where latency, accuracy, and reliability all had to hold at once. Drove the backend from a monolith into a monolith plus microservices plus serverless, over an event-driven layer, to keep service stable. An observability stack with Prometheus and Grafana, built from scratch. A datastore migration across 100K+ accounts with no downtime, running both databases live instead of taking a batch window. The subscription and enterprise licensing backend behind customers like Disney and Nike. Penetration test remediation three years running across four teams. On call for all of it. None of this was a fresh start. Everything was already running, with real users on it, and it had to keep running while I changed it.

How I work: most of what I've built started as a vague idea, and my job was making it specific enough for other teams to build against. I own my scope end to end, and once people start building, the spec always changes. That back-and-forth is where the design actually gets figured out.

Two projects I'm working on now:

Clearway: audits websites for accessibility and drafts the compliance report a specialist would write by hand. False alarms 43% to 23%, real issues caught 74% to 83%. My first AI reviewer scored well on my own gold set and near zero on the external W3C one, because it shared a rubric with the drafter. A second reviewer reading independently now catches 53% of those errors. Small-sample measurements, not proof.

VOX Personalis: fine-tuned Whisper with LoRA on my own Deaf-accented speech, because no off-the-shelf ASR handles it. WER 238.69% to 34.05%, now with sub-second streaming output. Next is sub-20% and on-device.

Communication style: I'm Deaf and work best with text-first collaboration, clear written specs, async updates, and captions/transcripts for meetings. This is how I've worked for 11 years, including with enterprise customers.

Looking for: Senior Backend IC, MTS, or Founding Engineer scoped to backend and agent infrastructure. Seed to Series B, applied AI, SF onsite or hybrid, full-time. Going deeper, not wider. I want to put a deep backend foundation to work on the part that’s hard: making a service reliable on top of an LLM that isn’t deterministic.