CHULIYING

Frontend-Focused Full-Stack Engineer

Chu Liying
emailgithub

ABOUT

AI APPLICATION INTEGRATION

I'm a frontend-focused full-stack engineer with a design background who builds AI products. I built a CRM with an AI assistant for a travel agency on my own, and co-develop the Rust agent runtime behind my company's AI assistant.

I care about the parts of an AI product people have to trust: where an answer came from, what an error means, and what happens when the model cannot give a supported answer.

CAPABILITIES

AI product interfaces
React and Next.js interfaces that show sources, loading, errors, and refusals, so users can tell what the system actually knows.
Retrieval & grounding
Exact facts come from SQL; retrieval is reserved for questions over approved aggregate text.
Response verification
Generated claims are checked against cited evidence, with a deterministic fallback when they fail.
Developer tooling & adoption
I drive adoption of AI-assisted development across a product team of about 35 (not direct reports), spanning engineering, product, design, and QA.

EXPERIENCE

HDRenewablesR&D Assistant Manager / Frontend Lead2024–Present
  • -Lead the frontend team for a multi-module energy platform; designed its configuration-driven routing and layout system with the backend team, cutting a new dashboard page from about two weeks to two days
  • -Built the platform's AI assistant end to end: identity forwarding, one streaming path, and sanitized model output
  • -Co-develop the Rust agent runtime behind it; authored its input-processing and per-user permission layers
  • -Line-manage four frontend engineers; lead planning, task allocation, and code review, and stay hands-on in production debugging
  • -Drive adoption of AI-assisted development across a product team of about 35 (not direct reports)
  • -Built a shared quality kit, added as a Git submodule, that keeps every route including logged-in pages under Lighthouse and gates pull requests on performance and accessibility baselines
EcosTekSenior Frontend Engineer2023–2024
  • -Built and maintained storefront and admin workflows for around 100 client accounts, integrating payment, logistics, and inventory systems
  • -Developed subscriptions, promotions, preorders, and a Shopee migration dashboard; added Cypress checks for promotion rules
  • -Planned frontend sprints, assigned work, and reviewed code
Brewing ArtsFull-Stack Engineer2021–2023
  • -Took the platform from concept to launch as its sole product designer and full-stack engineer
  • -Designed the content schema in Hygraph, a headless CMS, and built the Next.js frontend on its GraphQL API, with static generation and SEO for editorial content
  • -Set up hosting and deployment on Google Cloud
Independent ProjectsFreelance Product Designer / Frontend Engineer2020–2021
  • -Independently designed and built websites for clients, from the first layouts to the production frontend
The News LensFrontend Engineer2015–2020
  • -Worked on product planning, UX, and frontend development for a high-traffic bilingual media platform, including branded editorial sites and native advertising
  • -Independently took new brand sites from product planning and visual design through frontend implementation

EDUCATION

Ming Chuan University
Commercial Design · 2012

AI ENGINEERING PROJECTS

Travel Agency CRM & AI Assistant

Side project · In production

Independently designed and built a Customer 360 CRM and AI assistant for my family's travel agency, now in production on its real customer data. The requirement was trust: numbers from the database, claims backed by evidence, customer data kept inside.

Live demo ↗Architecture →Evaluation report →

  • Routed by data type: SQL answers counts and statuses, and only approved aggregate data goes to hybrid retrieval inside PostgreSQL, under the same row-level security as the records.
  • Numbers are computed by approved queries and text claims stay tied to cited evidence; invalid model output falls back to a deterministic answer. Customer-level data stays out of external-model calls.
  • Used 66 internal regression questions and seven governance checks, plus a separate 198-question paraphrase set for robustness. A thin state machine instead of an orchestration framework keeps every step testable.

Outcome

On 198 hidden paraphrases, 93.4% routed correctly and 89.4% passed end to end, with zero governance failures (synthetic data, one engine version); a de-identification gap meant the external-model path stayed closed. The read-only demo uses synthetic data and a local mock model.

  • Next.js
  • React
  • TypeScript
  • Python
  • PostgreSQL/pgvector
  • Cloudflare Workers AI
  • RAG

Rust Agent Runtime

Co-developed the Rust agent runtime behind the energy platform's AI assistant. My contributions: I authored the input-processing, authorization, streaming, and evaluation layers around the base agent my co-developer built.

  • One deterministic input layer in front of every model call, shared by REST, SSE, and the OpenAI-compatible endpoint with usage reporting and error handling, so the paths cannot drift apart.
  • Per-user identity instead of a shared token: the runtime fails closed when it cannot resolve permissions and narrows tools before the model sees them.
  • Chose degraded reports over refusals: partial access yields a report that names what it left out, and stored history is re-checked so revoked topics cannot leak back.

Outcome

Serves the energy platform's assistant with per-user permissions; 351 passing tests at the last snapshot. Accuracy and latency are not yet measured.

  • Rust
  • MCP
  • REST/SSE
  • OpenAI-compatible API
  • GitHub Actions

skill-commons

I created skill-commons, an open-source workflow kit for AI-assisted development. It targets spec drift: requirements lose detail as work moves on, and completion claims lose the check that would prove them.

  • Git is the state store: plans, decisions, and evidence live with the code, and delivery advances only with typed evidence such as a merge SHA.
  • Deterministic checks over model judgment: completion is a state a re-runnable check accepted, not a claim a report made.
  • One canonical source generates adapters for Claude Code, Codex, and Cursor, and adapted skills stay pinned to their upstream commit.

Outcome

Used in production work repositories, with twelve releases since July 2026. Its own evaluation found the cheapest routing mode scored lowest, so no token savings are claimed.

  • Python
  • Git
  • Claude Code
  • Codex
  • Cursor
  • CI/CD