I'm Jimmy (Jinrui) Li — a full-stack engineer based in Perth, working at the intersection of AI applications, backend systems, and product delivery.

Six years of backend engineering (Go / gRPC / Kubernetes at ByteDance and Zao'an, scaled to 500k DAU) gave me the discipline of distributed systems. The last two years — through my Master of IT at UWA and solo full-stack projects — I've moved that discipline into shipping LLM-powered products end-to-end: from OpenAI / Claude integration to streaming UIs to evals.

• ⚔️ Edge: backend rigour (microservices, observability, FinOps) + modern AI-app craft (GPT-4o, RAG, streaming, agentic workflows).
• 🎯 Now: building AI assessment & reporting platforms, an AI-driven mindfulness app, and a community cultural archive — all solo or small-team, all in production.

Stack: OpenAI GPT-4o · Claude · Nova AI · Prompt Engineering · Agent Prototyping

Integrating LLMs into real product flows: personalised reflection generation, narrative synthesis, contextualised insights, automated reporting. Streaming responses (SSE), prompt caching, and cost-aware design are non-negotiable.

Stack: React · TypeScript · Node.js · Express · Tailwind · shadcn/ui · Framer Motion

Solo-shipping production web apps end-to-end — auth, onboarding, interactive data flows, animations, accessibility (WCAG). Recent work: Soulstice Mind, WestPride Cultural Archive, AI Leadership Intelligence Platform.

Stack: Go · Python · gRPC · PostgreSQL · Neon Serverless · Drizzle ORM · Recharts

Built and scaled backend services at ByteDance (multi-tenant SaaS, billing & metering, user behaviour warehouse) and Zao'an (HuaYin IM system, 500k DAU). Latency, reliability, and observability are muscle memory.

Stack: Docker · Kubernetes · AWS · Azure · GCP · GitHub Actions · Prometheus · Grafana · Jaeger

Cloud-native deployments across the big three. Set up the Prometheus / Grafana / Jaeger stack that cut P95 latency to 500ms. CI/CD via GitHub Actions, with FinOps awareness from B2B SaaS work.

Stack: Eval-Driven Iteration · Zapier · Make · Lightweight Dashboards

I treat LLM features as probabilistic programs — test against a golden set, watch traces, iterate. No-code tooling (Zapier / Make) used when speed > custom code.

I don't chase "demo magic". I build for the boring path: streaming UX, prompt caching for cost, fallback on model failure, traces for every call. If non-technical stakeholders can't use it tomorrow without me, it's not shipped.

Six years of microservices taught me: observability before features. Every LLM call is a network call to a flaky third party — I instrument it the same way I'd instrument a payments API at ByteDance.

Most of my recent projects (Soulstice Mind, WestPride, AI Leadership Platform) are solo or near-solo builds. That forces brutal prioritisation: which feature actually moves the user forward, which is just craft satisfaction.