Xiayang Zhang

Experience

Independent AI Engineer — Self-directed R&D · Apr 2025 – Present · Melbourne

Building LLM/agent systems and operating fully AI-native.

• AREAR — live AI + AR education platform (primary engineer; a research collaboration with a university Research Fellow — live, on real research data — that began when the Fellow sought out my Dreemar AR-analytics work). A WebXR platform delivering: an agentic voice tutor with persistent cross-session memory; assessment via tool-call quizzes/games and per-lesson probes; personalization via per-student persona modelling; and feedback/analytics from audio emotion analysis plus spatial replay of student view-direction and movement.
  • Stack: Next.js 16 / React 19 / TS · Babylon.js + Three.js · multi-model — GPT-5.5 (analysis), OpenAI Realtime (voice), Gemini (dynamic storybook imagery) · Supabase/Postgres · Playwright + Vitest. (Private repo — live walkthrough on request.)
• Breadcrumbs — two-sided clinical-AI system (solo): Gretel, a real consumer iOS/Watch app live on the App Store (SwiftUI, on-device AI; non-diagnostic) — biosignal shifts → lightweight voice check-ins → structured, medication-aware, auditable episodes — and Hansel, a hospital-facing clinician console closing the loop — episode monitoring → AI-drafted follow-up suggestions (draft-only) → clinic check-ins pushed back into the app (React / TanStack; fully functional, shown as an all-synthetic demo — heavily regulated domain; demo sign-in at hansel.lemomo.ai) — integrating over a frozen, machine-checkable contract. Showcases correctness-first architecture — deterministic clinical executors (schema / terminology / medication checks) kept separate from a draft-only LLM adviser, with a provenance hub labelling every output real / mock / license-required / production-path. Cleared App Store health-data privacy review. gretel.lemomo.ai
• Harness-led / agent-native ops — I don't hand-write deep infra; I design the architecture and write the agent harness that drives agents to run it. Proof: /incus, a single-host IncusOS homelab (dual-Xeon / ZFS) running ~20 services and several long-running agent VMs, operated 24/7 by agents via an Agent Handbook + skills — plus a self-built multi-provider LLM gateway (GPT / Gemini / Grok / Claude) and multi-machine deployment. (Single-host homelab — not production multi-tenant scale.)
• AI-native operator — drive a fleet of models across my own dev/agent work — ~10.7B tokens/mo through Claude Code (~97% cache-hit) plus ~14.4B tokens/mo across self-run harnesses (incl. Codex, OpenCode), ≈$23K/mo equivalent API cost (my own dev/agent consumption, not product traffic — a conservative floor).
• Open-source (github.com/2nd1st) — AIMA, an early cognitive-agent framework (TypeScript · Bun · Postgres/pgvector); api-log, an LLM-gateway recording proxy (Go); claudoros, a Claude Code focus monitor (Python). Also building: api-log-analytic (Go / SQLite / Svelte — large-scale LLM-traffic analysis & ingestion correctness) and loomomo (Rust — a baseline-delta CI gate that flags only the findings an agent's change introduced). (work samples; releasing.)

AR/AI Software Developer & Multimedia Specialist — Dreemar · Nov 2021 – Mar 2025 · Melbourne

Sole developer for the final ~1.5 years — effectively the entire engineering team: built, updated and maintained the frontend, backend, CMS, web editor and mobile apps single-handed, plus all production AWS operations (shipped a full app solo in year one). Worked in an Agile team for the first ~2 years before becoming sole developer.

• Built ARI-AI — an early-stage multi-agent AI marketing-content SaaS (pre-users), solo, end-to-end, deployed on AWS: Next.js + TypeScript frontend (NextAuth, Prisma); a Python CrewAI multi-agent backend; a LiteLLM gateway across OpenAI + Anthropic with a deliberate Claude-for-orchestration + GPT-for-content split; and full observability (Langfuse tracing/cost, Prometheus) on a Dockerised multi-service stack with CI/CD.
• Shipped 6+ AR apps to the App Store / Google Play (Unity / C#) — mostly kids' education AR, including a Deakin University collaboration and Victorian Tech School apps; Dreemar's broader app portfolio has tens of thousands of downloads. Also DreemXR (React Native) — a downloadless cross-platform marker-tracking AR app (iOS App Clip + Android Instant App).
• Maintained the AR Online Editor / CMS (Angular · .NET · Unity-WebGL editor); built immersive AR/3D experiences and the platform's spatial analytics — position/orientation capture for session replay, attention targets, popular paths, and dwell-time patterns; produced multimedia/design assets.

Earlier — Game Developer (self-employed, Unity / C#, 2020–21) · Digital Designer intern @ Triangle Financial (web + AWS, 2020) · Graphic Designer intern @ MountainTop Education (2019).