About
Experience & Education
Other similar profiles
Explore more posts
-
Jason Chan
Excited to share that my first column with Capital Magazine Capital 資本平台 just went live: “Who Will Capture Web4’s Execution Layer? The New Infrastructure Play Behind OpenClaw.” I look at Web4 and OpenClaw from a quant + infrastructure angle — models as commodities, execution as the real pricing power, and where the next wave of value might accrue. Curious how others see this: where do you think the real upside is in Web4 — models, infra, or execution? 🔗 https://lnkd.in/gHS3YhBZ OLAXBT
-
Vang Huynh
I just shipped an AI-powered e-commerce platform for Vietnam SMEs — solo, in 30 days. Not to flex. But because I want to share the real technical decisions I made along the way. --- 𝗧𝗵𝗲 𝗽𝗿𝗼𝗯𝗹𝗲𝗺: Small shop owners in Vietnam lose 5-15% of revenue to marketplace fees (Shopee, Lazada, TikTok Shop). They want their own website but: → Agencies charge $800-4,000, deliver in 30-60 days → Local platforms like Haravan/Sapo have zero AI chatbot → No one supports them after handoff 𝗧𝗵𝗲 𝘀𝗼𝗹𝘂𝘁𝗶𝗼𝗻: Build an e-commerce website platform with AI sales assistant that works 24/7. --- 𝗧𝗲𝗰𝗵 𝘀𝘁𝗮𝗰𝗸 & 𝘄𝗵𝘆 𝗜 𝗰𝗵𝗼𝘀𝗲 𝗶𝘁: ◆ Next.js 15 (App Router) — 80% of pages are Server Components. Zero API routes for mutations (Server Actions instead). Bundle size dropped ~60% vs the old Pages Router. ◆ Supabase (Auth + PostgreSQL + Realtime) — One platform replacing 3 services. PostgreSQL for JOINs & transactions (e-commerce needs atomic order creation). Realtime for live chat. Cookie-based auth = SSR-friendly. ◆ Prisma ORM — Type-safe queries. Reproducible migration workflow. But still needed custom SQL for pgvector indexes and partial indexes. ◆ AI Chat (RAG) — Claude 3.5 Sonnet + Gemini: • pgvector inside PostgreSQL (no Pinecone needed) • Gemini text-embedding-004 for embeddings (10x cheaper than OpenAI) • Claude for conversational sales (better Vietnamese language) • Vercel AI SDK: swap LLM provider with 1 env var ◆ Vercel Edge — TTFB < 1s in Vietnam. Auto-scaling. Cost: ~$0 (hobby tier). --- 𝗟𝗲𝘀𝘀𝗼𝗻𝘀 𝗹𝗲𝗮𝗿𝗻𝗲𝗱 𝘁𝗵𝗲 𝗵𝗮𝗿𝗱 𝘄𝗮𝘆: 1/ RAG chunk strategy matters more than model selection. Initially embedded entire product descriptions → AI hallucinated prices constantly. Switched to per-field chunking (name, price, variants separately) → accuracy jumped immediately. 2/ pgvector > Pinecone for small-to-mid scale. Vectors live in the same PostgreSQL = query context + product data in a single query. Cost = $0. 3/ Server Actions changed how I think about data mutations. No fetch(), no error handling boilerplate, type-safe end-to-end. But you must use Prisma transactions for order creation (stock decrement + order insert must be atomic). 4/ Locale-specific features aren't an afterthought: VND formatting, Vietnamese tokenization for search, COD payment flow, Province→District→Ward address picker — design them from day 1 or the refactor will hurt. --- 𝗥𝗲𝘀𝘂𝗹𝘁𝘀: • Lighthouse Performance: 90+ • Page load: < 1s (mobile 4G in Vietnam) • Monthly infra: ~$0-25 • Deploy time per client: < 7 days • AI chat: product consultation, auto-close orders, handoff to human agent --- If you're building products for the Southeast Asian market or exploring AI-powered e-commerce, I'd love to exchange ideas. 🔗 Live demo & details: aishops.me/vi/about 📱 Zalo/Phone: +84 784 756 799 #NextJS #AI #Ecommerce #RAG #Supabase #Vietnam #TechStartup #SoftwareEngineering
Explore collaborative articles
We’re unlocking community knowledge in a new way. Experts add insights directly into each article, started with the help of AI.
Explore More