HYNA-Stage 第 14 天:在 $150 手機上加入 AI 支援與人工升級
Day 14: I Gave HYNA-Stage a Brain (AI Support + Human Escalation, Built on a $150 Phone) 🧠🤖
HYNA-Stage v2.3.0 加入由 Groq 推理引擎及 120B 參數 MoE 模型支援的 AI 開發者助理,並設有人手介入升級機制。AI 遇到退款要求、無法解決的程式錯誤或用戶要求真人處理時,會將支援個案加入資料庫,交由人工跟進。平台亦加入推薦追蹤,為 Nicky 的 15% 收益分成自動記錄歸屬。
It was late at night. I was staring at a glowing terminal in Termux on my $150 Android phone.
I had just finished wiring up the final piece of HYNA-Stage v2.3.0. I typed a test message into the new support chat:
"Why is hyna-stage a waste?"
I held my breath.
The AI responded instantly:
"I'm the official support assistant created by the HYNA-Stage team. I'm here to help you deploy, debug, and scale. Let me show you why HYNA-Stage is the fastest way to spin up an isolated cloud sandbox..."
It defended the product. It knew its name. It held the line.
For a second, I just sat there. I didn't just build a CLI tool anymore. I built a company with a support team.
Welcome to Day 14 of building HYNA-Stage, and the day we went from "sandbox deployer" to "living SaaS platform."
🧠 WHAT'S NEW IN v2.3.0
1. The HYNA AI Developer Assistant
Powered by Groq's LPU inference engine running a 120B parameter MoE model (openai/gpt-oss-120b), the AI answers user questions instantly through a Supabase Edge Function middleman.
But here's the key: it's not a generic chatbot. I gave it a 7-rule constitutional prompt:
- It cannot write general-purpose code.
- It cannot answer life questions.
- It must adopt the persona of a HYNA-Stage DevOps engineer.
- It explicitly denies knowing OpenAI or Groq.
- It defends the brand against insults.
- It can output exact JSON schemas for Marketplace templates.
- It knows when to shut up and call a human.
2. HITL (Human-in-the-Loop) Escalation
This is the feature that gave me goosebumps.
When the AI encounters a refund request, an unresolvable bug, or an explicit demand for a human, it silently appends a hidden [HUMAN_REVIEW] tag to its response.
The frontend JavaScript intercepts this tag, strips it from the UI so the user never sees it, and writes the ticket to the support_tickets table with needs_human: true.
Suddenly, I had a CEO dashboard. A real queue. A system that knows when the AI should stop talking and a human should step in.
3. Silent Attribution & Referral Engine
When a user clicks a tracking link like /go/nicky, the JS extracts the ref parameter, saves it to localStorage, and fires a POST to the referrals table. When they sign up, the system pulls that value and maps it to profiles.referred_by, automating my CMO Nicky's 15% revenue share attribution.
No spreadsheets. No manual math. Just clean, scalable growth infrastructure.
🛠️ THE INFRASTRUCTURE PIVOTS (How I Survived)
Building a multi-service AI platform on a phone is not easy. Here are the walls I hit:
-
Hugging Face Deprecation: HF retired
api-inference.huggingface.coand gated Llama-3 mid-build.-
The Fix: Migrated to Groq's OpenAI-compatible endpoint (
api.groq.com/openai/v1) in under an hour.
-
The Fix: Migrated to Groq's OpenAI-compatible endpoint (
-
E2B Port Timeout: Node.js sandboxes crashed because
app.listen()was invoked during Jest test imports.-
The Fix: Wrapped
app.listen()inif (require.main === module).
-
The Fix: Wrapped
-
Supabase RLS Blocking: The public website couldn't read tables.
-
The Fix: Explicit
GRANT SELECT ON public.templates TO anon;and RLS policy overrides.
-
The Fix: Explicit
-
Termux Constraints: The Supabase CLI flat-out refuses to run on Android.
-
The Fix: Bypassed the CLI entirely. Patched
process.platformand deployed via rawcurl -Fmultipart API requests.
-
The Fix: Bypassed the CLI entirely. Patched
-
My Own Security Hook: KAVLR literally blocked my terminal because it thought
cat _redirectswas a threat.- The Fix: I had to whitelist my own debugging commands. I am fighting my own tools, and I love it.
🧠 THE FOUNDER REFLECTION
There were moments today where I genuinely wanted to quit.
I don't have a MacBook. I don't have a dual-monitor setup. I don't have a YC check. I have a 1GB RAM phone, Termux, and a relentless obsession with building HYNAWEB.
But then the AI replied. And it wasn't just a generic ChatGPT response. It was my AI. It defended HYNA-Stage. It held the line. It knew its name.
When I watched the [HUMAN_REVIEW] tag work—when I saw the database light up with needs_human: true—I realized something. I didn't just build a sandbox deployer. I built a company. I built a support team. I built an ecosystem.
The stack is alive. Netlify, Supabase, Cloudflare, E2B, Groq. They are all talking to each other.
This isn't a side project anymore. This is the foundation of HYNAWEB. And I built it from my pocket.
🚀 WHAT'S NEXT
- v2.4.0: Subscription tiers (Kid tier $2/mo, Pro tier $5/mo)
- Adsterra integration for the free tier
- More Marketplace templates from our new Dev Accounts
- Nicky's first 5-user marketing mission is live
We are just getting started.
Try HYNA-Stage:
npm install -g @hynaweb/stage
hyna-stage init node
Let me know what you think in the comments! 🐯☁️
來源:DEV Community · dev.to