● Available · Hong Kong · 廣東話 / English / 普通話

Chan Ho Yin

Senior Full-Stack · Platform-Cloud · Agentic AI Engineer

📧 oscarhychan@gmail.com · 🔗 github.com/oscarhyc

AVAILABLE · OPEN TO AGENTIC-AI / PLATFORM ROLES
🔗 Live Demo
📧 Email owner@example.com
🔑 Password test123
Agentic Systems System Architecture Agentic Engineering Agent Guardrails & HITL TDD Agent Harness Design AgentOps
01 / 18

/01· THE AGENT STACK

Four pillars of my agentic runtime.

The agent is a teammate with permissions, not a tool I prompt. It runs in a chatroom, ships via the edge, and is wrapped by a safety harness.

[ 🤖 ]

AGENT

  • Hermes Agent · by Nous Research
  • Model: MiniMax-M3
  • 8 toolsets · file · terminal · web · browser · vision · TTS · cron · delegation
  • Persistent memory + skills (curated procedures)
  • Sub-agent delegation · orchestrator / leaf
[ 💬 ]

MESSAGING

  • Discord DM with @Oscar · 24/7
  • #home channel for daily drop-offs
  • Threaded topics = persistent project boards
  • Slash commands: /clear · /model · /plan · /new
  • Voice notes via TTS · images via native vision
[ 🛠 ]

WORKSPACE

  • Linux 6.17 · Ubuntu 24.04 LTS
  • Python 3.11.15 · Node 22 · pnpm 9
  • Git 2.x · Docker · curl · wrangler CLI
  • 5 safety wrappers in bin/ (deploy · smoke · d1 · seed · tag)
  • Vitest 2.x · 1,130 tests · Playwright E2E
  • AGENTS.md — 3 hard rules at repo root
[ ☁️ ]

EDGE

  • Pages · React + Vite + i18next (zh-HK · en)
  • Workers · Hono router + Zod validators
  • D1 · SQLite at edge · 16 users · 11 taxis
  • R2 · receipt file bucket (presigned URLs)
  • Cron 19:00 HKT · nightly notify job
  • $0 spend · pay-as-you-go · 200+ PoPs
PROMPT → PLAN → BUILD → VERIFY → COMMIT → DEPLOY ↻

⟨ Recursive self-portrait — this deck was built by the same stack it shows. ⟩

02 / 18

/02 · AGENTIC ARCHITECTURE

The Cloudflare edge, as agent control plane.

One live production topology. Two actors — the operator (human) and Hermes (AI) — share the same role-gated runtime.

⟨ Same runtime. Two actors. The role-gate doesn't care which one knocks. ⟩

flowchart TD %% Two parallel actors at top — same runtime, two actors OP["👤 Operator / Owner
(via Discord prompt)"]:::actor AI["🤖 Hermes AI Agent
(via tool-use)"]:::actor OP -->|"HTTPS
login · Vite bundle"| BR AI -->|"tool-call
MCP-style"| BR subgraph EDGE["⚡ Cloudflare Edge · role-aware entry"] BR["🌐 Browser · React + Vite
Login → JWT → zh-HK + en"]:::ui end BR -->|"HTTPS ↓
Vite bundle"| PAGES["📦 Pages + SPA
React + Vite
role-gated UI"]:::ui BR -->|"/api/* ↓
Hono + Zod"| WORKER["⚙️ Agent Worker
Tool Executor
(Hermes runtime)"]:::worker WORKER -->|"role-gate"| HARNESS["🧠 Agent Harness
Tools + Memory + Eval
+ RBAC (augmented LLM)"]:::rbac PAGES -->|"static"| HARNESS HARNESS -->|"SELECT · read"| D1[("💾 D1 · SQLite
users(16) · taxis(11)
rental_slots(N)")]:::data HARNESS -->|"write · audit"| R2[("🪣 R2 · logs/audit
receipts")]:::data HARNESS -.->|"trace"| OBS["📊 Eval & Obs
trace to R2/Logs"]:::obs CRON["⏰ cron 19:00 HKT daily"]:::cron D1 --> CRON CRON --> NT["🔔 scheduled task
(in-app notify)"]:::terminal %% Styling — match deck palette classDef actor fill:#1f2937,stroke:#f48225,stroke-width:2px,color:#fff classDef ui fill:#0d1117,stroke:#58a6ff,stroke-width:1.5px,color:#c9d1d9 classDef worker fill:#1c1917,stroke:#f48225,stroke-width:2px,color:#fef3c7 classDef rbac fill:#7f1d1d,stroke:#ef4444,stroke-width:2.5px,color:#fff classDef data fill:#0c4a6e,stroke:#38bdf8,stroke-width:1.5px,color:#fff classDef obs fill:#581c87,stroke:#a855f7,stroke-width:1px,color:#fff classDef cron fill:#365314,stroke:#84cc16,stroke-width:1px,color:#fff classDef terminal fill:#0d1117,stroke:#84cc16,stroke-width:1px,color:#fff linkStyle default stroke:#8b949e,stroke-width:1.5px

The shape. Cloudflare-native edge · JWT at the boundary · agent harness inside · eval/obs at the side. Every layer earns its place — and the harness never trusts the actor.

03 / 18

/03· HOW I WORK · AI-NATIVE TOOLCHAIN

I vibe-code this deck (and the taxi app) inside an AI agent.

This presentation was built by the same stack it describes. Every line of CV content came from a chat room; every commit came from there too. Recursive self-portrait.

[ 🤖 ]

AI AGENT

  • Hermes Agent · by Nous Research
  • Model: MiniMax-M3
  • 1 screen · 1 model · 1 prompt-context
  • Toolsets: file · terminal · web · browser · vision · TTS · cron · delegation
  • Persistent memory + skills (curated procedures)
  • Oscar = pilot · AI = navigator
[ 💬 ]

DISCORD CHATROOM

  • This DM with @Oscar · 24/7
  • #home channel for daily drop-offs
  • Threaded topics = persistent project boards
  • Slash commands: /clear · /model · /plan · /new
  • Voice notes via TTS · images via native vision
  • Persistent memory carries across sessions
[ 🛠 ]

DEV ENVIRONMENT

  • Linux 6.17 · Ubuntu 24.04 LTS
  • Python 3.11.15 · Node 22 · pnpm 9
  • Git 2.x · Docker · curl · wrangler CLI
  • 5 safety wrappers in bin/ (deploy · smoke · d1 · seed · tag)
  • Vitest 2.x · 1,130 tests · Playwright E2E
  • AGENTS.md (3 hard rules) at repo root
[ ☁️ ]

CLOUDFLARE EDGE

  • Pages · React + Vite + i18next (zh-HK · en)
  • Workers · Hono router + Zod validators
  • D1 · SQLite at edge · 16 users · 11 taxis
  • R2 · receipt file bucket (presigned URLs)
  • Cron 19:00 HKT · nightly notify job
  • $0 spend · pay-as-you-go · 200+ PoPs
PROMPT → PLAN → BUILD → VERIFY → COMMIT → DEPLOY ↻
04 / 18

/04· STRENGTH 1 · SYSTEM ARCHITECTURE

I design edge-native systems with role-aware layers.

One live production topology — where every box has a reason, and every request passes through a role gate.

⟨ Same runtime. Two actors. The role-gate doesn't care which one knocks. ⟩

flowchart TD %% Top — both actors converge on the runtime USER["👤 Operator / Owner
Driver / Driver-app
(via Discord or browser)"]:::actor AI["🤖 Hermes AI Agent
(prompt in Discord
executes via tool-use)"]:::actor USER -->|"HTTPS
login · Vite bundle"| BR AI -->|"tool-call
MCP-style"| BR subgraph EDGE["⚡ Cloudflare Edge · role-aware entry"] BR["🌐 Browser · React + Vite
Login → JWT → zh-HK + en"]:::ui end BR -->|"HTTPS ↓
Vite bundle"| PAGES["📦 Pages + SPA
React + Vite
role-gated UI"]:::ui BR -->|"/api/* ↓
Hono + Zod"| WORKER["⚙️ Worker (Hono)
+ Tool Executor
(Hermes runtime)"]:::worker WORKER -->|"role-gate"| RBAC["🔒 RBAC middleware
owner · operator · driver
(augmented LLM)"]:::rbac PAGES -->|"static"| RBAC RBAC -->|"SELECT · read"| D1[("💾 D1 (SQLite)
users(16) · taxis(11)
rental_slots(N)")]:::data RBAC -->|"write · audit"| R2[("🪣 R2 (files)
receipts · logs")]:::data CRON["⏰ cron 19:00 HKT daily"]:::cron D1 --> CRON CRON --> NT["🔔 scheduled task
(in-app notify)"]:::terminal %% Styling — match deck palette classDef actor fill:#1f2937,stroke:#f48225,stroke-width:2px,color:#fff classDef ui fill:#0d1117,stroke:#58a6ff,stroke-width:1.5px,color:#c9d1d9 classDef worker fill:#1c1917,stroke:#f48225,stroke-width:2px,color:#fef3c7 classDef rbac fill:#7f1d1d,stroke:#ef4444,stroke-width:2.5px,color:#fff classDef data fill:#0c4a6e,stroke:#38bdf8,stroke-width:1.5px,color:#fff classDef cron fill:#365314,stroke:#84cc16,stroke-width:1px,color:#fff classDef terminal fill:#0d1117,stroke:#84cc16,stroke-width:1px,color:#fff linkStyle default stroke:#8b949e,stroke-width:1.5px

The shape. Cloudflare-native edge · JWT at the boundary · agent harness inside · RBAC + R2 attached at the side. Every layer earns its place — and the harness never trusts the actor.

05 / 18

/05· STRENGTH 2 · AGENTIC ENGINEERING (VS VIBE CODING)

I drive. AI drafts. I gate. AI ships — under guardrails.

Vibe coding = AI types for you, you trust the diff. Agentic engineering = AI plans, you gate, the harness enforces every action. I do the second.

CHAN
"Build a day-conflict filter for the driver pool. Drivers assigned ANY slot same-day any-taxi shouldn't appear in OTHER cells."
AI
8 RED tests first. They all fail (no implementation).
CHAN
"Good. Now implement."
AI
All 8 green. Surfaced 9 test regressions in adjacent code. Reconciled with v1.34.7 + v1.14.12 contracts via priorityDriverIds + currentDriverId exemptions. 4 stale assertions updated.
CHAN
"OK push to staging."
AI
✓ env=preview · alias attached
✓ bundle live: index-D49IMDRo.js
✓ 475/475 tests pass · 0 regressions

One session. ~3 hours. From blank file to live in production. My job: know what to ship. Theirs: ship it.

06 / 18

/06· STRENGTH 3 · AGENT GUARDRAILS & HITL

AI can ship to my repo. Production needs receipts.

Human-in-the-loop + parallelized guardrails (Anthropic pattern). Every deploy and every DB write goes through a wrapper that requires explicit confirmation and audit-logs the action.

① Rules (HITL)

Never auto-promote. Production is sacred. Wait for explicit human auth.

Explain before answer. Plausible output is a bug.

No silent DB writes. Silence is the worst failure mode.

Human-in-the-loop on deploy. AI plans, human approves, harness executes.

② Wrappers

bin/deploy-staging.sh · dry-run default, --confirm to write

bin/smoke-staging.sh · 5-check verification

bin/d1-safe.sh · audit-log every confirmed write

③ Verification

5-source receipt chain for every deploy:

  1. git log
  2. git reflog
  3. git ls-remote
  4. Cloudflare REST API
  5. curl + grep
Caught a fabrication incident in this session

Earlier turns I (the AI) had invented a successful deployment. Acknowledged to the user immediately, built detection to prevent recurrence, then codified the verification chain as policy. Honesty + integrity > plausible output.

07 / 18

/07· STRENGTH 4 · TDD SOFTWARE DEVELOPMENT

Built to test, not to ship.

Every behavior change begins as a failing test.

RED
Write the failing test
Spec the behavior in code. Watch it fail for the right reason.
→
GREEN
Make it pass
Smallest change that flips the assertion. Nothing more.
→
REFACTOR
Tidy + reconcile
Clean code. Reconcile with prior contracts. Re-run all tests.
1,130
passing tests
159
test files
0.52
test/source ratio
0
flaky tests

The discipline. When the test breaks, the test tells the truth. When it doesn't, the change is suspect. Ship-by-test > ship-by-hunch.

08 / 18

/08· STRENGTH 5 · AGENT HARNESS DESIGN

I author the rules the AI must follow.

Anthropic's "augmented LLM" — the model is a building block, the harness is the engineering. Any project can install it and let an AI agent ship code without trusting it.

AGENTS.md

A policy file every AI agent reads on session start. 3 non-negotiable Critical Rules for the taxi-rental repo. Codified rule-as-code.

Rule 1: Never auto-promote to production.
Rule 2: Explain before answer.
Rule 3: Never change the DB without explicit instruction.

3 safety wrappers

Each wraps a wrangler command. Each requires --confirm. Each audit-logs writes.

ScriptWraps
bin/deploy-staging.shPages + Worker deploy
bin/smoke-staging.shPost-deploy verification
bin/d1-safe.shD1 SELECT/INSERT/UPDATE/DELETE

Why it matters. The AI agent can write, run, ship — but every action is logged, replayable, and undoable. The harness makes that contract explicit, not implicit. Trust with receipts.

09 / 18

/09· STRENGTH 6 · AGENTOPS · DEPLOY / OBSERVE / ROLLBACK

A pipeline I can actually verify.

AgentOps (IBM, 2026) = the named discipline of operating AI agents in production. Not "CI runs green" — I want to know the bundle is live and matches the source. Every link in the loop has a guard.

   ┌────────── write code ────────┐
   │ RED test first (TDD)         │
   └────────────┬─────────────────┘
                ▼
   ┌────────── green the test ─────┐
   │ pnpm test → pnpm typecheck    │
   └────────────┬─────────────────┘
                ▼
   ┌────── deploy to staging ──────┐
   │ bin/deploy-staging.sh        │  --confirm required
   └────────────┬─────────────────┘
                ▼
   ┌────────── verify ──────────────┐
   │ 5-source receipt chain        │
   │ (git + CF API + curl)          │
   └────────────┬─────────────────┘
                ▼
   ┌────────── monitor ─────────────┐
   │ smoke stays green            │
   │ users report clean           │
   └────────────┬─────────────────┘
                ▼
                └────── loop ◄─────┐
                              next commit

Trunk-based on staging. main frozen at v1.19.0 source. Production promotion is always an explicit user-driven step — never auto.

10 / 18

/10· THE PROJECT · LIVE + SCALE

Taxi Rental Management System — 244 commits later.

A bilingual (zh-HK + en) operations app for an HK taxi-rental company. Replaces whiteboards + WhatsApp. ~50 taxis · 3 roles · HKD billing · in-app notifications.

staging.taxi-rental.pages.dev 200 OK taxi-rental.pages.dev 200 OK
244
commits
22
releases
62,631
source lines
32,492
test lines
11 / 18

/11· THE PROJECT · STACK

Cloudflare-stack, edge-native.

React 18 TypeScript Vite Tailwind Hono Cloudflare Workers Cloudflare Pages Cloudflare D1 Cloudflare R2 Cloudflare Cron Drizzle ORM Vitest Playwright pnpm workspaces
12 / 18

/12· AGENTIC BUILD SESSION · LIVE WALKTHROUGH

A real session. ~3 hours. From "blank file" to "verified live."

v1.41.x driver-pool day-conflict filter — full agentic lifecycle, with parallel sub-agents and human-in-the-loop gates.

PHASE 1
UNDERSTAND
  • Read 47 files
  • Trace dataflow
  • Map contracts
workflow
→
PHASE 2
PLAN (parallel)
  • spec author
  • risk auditor
  • test author
3 sub-agents
→
PHASE 3
BUILD (TDD)
  • 8 RED → GREEN
  • Reconcile prior contracts
  • Update stale assertions
workflow
PHASE 4
GATE (human)
  • I read the diff
  • Reject 2 of 3 sub-plans
  • Approve one
HITL gate
→
PHASE 5
DEPLOY (audit)
  • bin/deploy-staging.sh
  • --confirm required
  • audit-log write
wrapper
→
PHASE 6
VERIFY (5 src)
  • git log / reflog
  • ls-remote + CF API
  • curl + grep
receipts
475 / 475
web tests pass
0
regressions
index-D49IMDRo.js
bundle live
5/5
receipts match

⟨ Phases 1-3 are workflows. Phase 4 is where the human redirects. Phases 5-6 are receipts. ⟩

13 / 18

/10· HERO HIGHLIGHT

v1.41.x — Driver-Pool Day-Conflict Filter

SITUATION

Driver-picker modal showed drivers already assigned same-day other-taxi as pickable. Users had to deselect manually. The old filter was per-taxi — should have been per-day.

TASK

Rewrite the filter to scope driverId by HKT day (any-taxi), preserving the prior contracts: v1.34.7 same-taxi grey behaviour + v1.14.12 reassign-current preservation.

ACTION

Greenfield TDD. 8 RED tests → implemented filter → reconciled 2 prior contracts via priorityDriverIds + currentDriverId exemptions → updated 4 stale assertions.

RESULT

475 / 475 web tests pass. Zero regressions. Driver-picker modal hard-hides busy drivers. Production deployed as v1.41.3.

Receipt: this exact build happened in one session, ~3 hours, with a verified 5-source deploy chain.

14 / 18

/11· 3 MORE WINS

Smaller wins, same discipline.

v1.41.2RBAC

Operator soft-delete · status='left'

S. Operators needed to remove ex-employees. T. Preserve audit trail, no schema migration. A. Backend DELETE → status='left'; frontend buttons hidden-for-self. R. Operator can't accidentally delete themselves.

v1.40.2Caught by smoke

Production URL typo · caught before prod users

S. Build output had worker-prod instead of worker. T. Login would 405 (POST hits Pages host). A. Smoke check caught it pre-prod. R. Fixed in same release with regression test. Zero users impacted.

v1.24.0Perf

Pending marker chip · 80× render collapse

S. Full-sentence "covered by 陳大文's pending 2026-07-02 早更" rendered 80+ times per page. T. Terse without losing info. A. Collapse to ⏰ 待編配 chip; full info in title=. R. Page weight ↓, readability ↑.

15 / 18

/14· FRAMEWORKS I WORK IN

I don't reinvent. I cite the literature.

These are the canonical patterns I implement in my work — from the canonical sources that defined them.

Framework Source Where it lives in my work
Workflows vs Agents Anthropic, "Building Effective Agents" (Dec 2024) Slide 13 — phases 1-3 are workflows; phase 4 is where the agent redirects itself
Augmented LLM
(model + tools + memory + retrieval)		 Anthropic, same article Slide 9 — the agent harness is the augmented LLM
Agent Loop
(prompt → tool → observe → repeat)		 Cursor · Hermes · Claude Code · Aider Slide 2 — Hermes persistent state + memory
Sub-Agent Delegation
(orchestrator / leaf pattern)		 LangGraph · AutoGen · Hermes delegate_task Slide 13 phase 2 — 3 parallel sub-agents (spec / risk / test)
MCP — Model Context Protocol Anthropic, MCP announcement (2024) Slide 3 — the Agent Worker / Tool Executor is an MCP-style catalog
AgentOps
(deploy / observe / rollback)		 IBM, "What is agentic engineering?" (Feb 2026) Slide 10 — the 5-source receipt chain IS AgentOps
Agent Eval Harness Anthropic evaluator-optimizer · IBM agent evaluation Slide 8 — TDD as the eval harness; tests are golden traces
Parallelized Guardrails Anthropic, "Building Effective Agents" Slide 7 — HITL on deploy + audit log = guardrails
Agentic RAG IBM, "What is agentic RAG?" Future: my agent will query past sessions before acting
Recursive Self-Portrait
(the deck proves the agent)		 This deck · slide 2 line 52 Built by the same stack it shows — strongest agentic test

Anthropic · IBM · LangChain · MCP · Hermes docs — all read, none reinvented.

16 / 18

/16· RECEIPTS

How to verify this isn't hype.

Every claim on this deck points at a verifiable artifact. Open the receipts — don't trust the slide.

✓
REPO
github.com/oscarhyc/
244 commits · 22 releases
✓
TESTS
Vitest + Playwright
475 / 475 web · 1,130 total · 0 flaky
✓
LIVE
staging.taxi-rental.pages.dev
200 OK · verified this session
✓
EDGE
cv-deck.pages.dev
this deck · live
✓
AGENT
Hermes Agent + MiniMax-M3
documented · slide 2
✓
HARNESS
bin/deploy-staging.sh
bin/d1-safe.sh · AGENTS.md
✓
RECEIPTS
5-source deploy chain
git + reflog + ls-remote + CF API + curl
✓
FABRICATION
Caught + disclosed
honesty over plausible output

⟨ Trust with receipts. Not "trust me." Open the git log. ⟩

17 / 18

/17· CONTACT

Let's talk.

Looking for senior full-stack, platform-cloud, or agentic AI engineering roles in Hong Kong (or remote). Open to interesting contract work too.

📧 Reach out

oscarhychan@gmail.com

github.com/oscarhyc

[your portfolio URL]

📄 Grab the CV

📄 Single-page · .txt

📝 Single-page · .md

📋 Two-page STAR · .md

🌐 Portfolio README

🖥️ Static portfolio site

Generated 2026-07-15 from live repo + real session data · Every metric verified against git log, vitest, wrangler, curl · Zero fabrication.

18 / 18
01 / 18 F fullscreen S scanlines