Why I Chose T3 + Supabase + Vercel
A 51-year-old non-developer used semiconductor-grade methodology to pick a tech stack — and here's the full record

1. Why I Took Tech Selection Seriously

I'm not an engineer. I'm a 23-year B2B sales professional. Before building the Coastline platform, my knowledge of "T3 Stack" was zero. Same for "tRPC."

But I know one thing: the wrong infrastructure choice costs far more than the technology itself.

I've watched too many early-stage founders choose a "cool" tech stack and discover six months later that it can't support their core feature set — forcing a full rewrite. That's not just a technical problem. It's a time problem, a money problem, a market timing problem.

So when I was choosing the technical foundation for the Coastline platform (mycoastline.ai), I didn't ask "which is most popular." I asked: Which option lets a non-coder use AI agents to build a CRM + Sales Pipeline + Community + Shop platform that won't need re-architecting for 3 years?

The answer took two days of AI agent collaboration to produce. This post is the complete record of that decision — not just "what I chose," but "why I chose it, and why the others weren't good enough."

2. SEMI E10-0304: Semiconductor-Grade Evaluation Methodology

In the semiconductor industry, selecting an equipment vendor doesn't rely on gut feeling or sales relationships. It relies on a rigorous Vendor Qualification Document (VQD) methodology.

My AI Technical Architect, Mike, adapted this framework to cloud infrastructure evaluation using WMCDA (Weighted Multi-Criteria Decision Analysis):

All 12 criteria have explicit scoring rubrics (1-10), scored independently by 5 AI judges, then weighted and averaged. I intentionally gave "Developer Experience" the highest weight (15%), because AI agent development velocity directly determines how quickly ideas become products.

3. Four Tech Stack Options

My original choice was Stack 1 — Vercel + Railway + Fly.io. That was the decision on February 27, 2026. In early March, I revisited it, found it "too hasty," and commissioned this full VQD evaluation. Four stacks were assessed:

Stack 1: Vercel + Railway + Fly.io (My Original Choice)

This was my initial instinct. Three well-known services with clear roles: Vercel for frontend, Railway for backend and database, Fly.io for containers and WebSocket. It looks reasonable — but the problem is, managing 3 vendors = 3 bills + 3 deployment pipelines + 3 failure points. Railway also suffered a major outage in January 2026, exposing infrastructure maturity issues. Score: 7.18. 3-year cost: $5,160.

Stack 2: Cloudflare Full Stack (Cheapest, but Fatal Flaws)

Cloudflare is highly attractive — single vendor, 330+ global PoPs (including Taipei), 3-year cost just $1,800. But it has two dealbreakers I couldn't accept:

• D1 database is SQLite, not PostgreSQL — cannot support Row-Level Security (RLS), which is the core of Coastline's 3-tier data isolation for CRM
• Cloudflare Access has no native LINE Login support — Taiwan has 466 LINE community members who need LINE OAuth to sign in

Both are hard technical dealbreakers. Score: 6.12, last place.

Stack 3: Supabase + Vercel (Very Good, But Missing One Layer)

Stack 3 and Stack 4 use identical infrastructure (Supabase + Vercel). The only difference: Stack 3 lacks Turborepo Monorepo and tRPC. Without tRPC, frontend and backend types are separate — requiring manual synchronization. For AI-agent development, this is a significant quality risk. Score: 7.84. 3-year cost: $3,900.

4. Key Comparisons: Database, Auth, Deployment

Among the 12 evaluation criteria, three decisive differences determined the winner:

Database: PostgreSQL vs. SQLite

Coastline CRM requires 11 tables, complex JOIN queries, Row-Level Security (RLS), and full-text search. PostgreSQL supports all of these natively; Cloudflare D1 (SQLite) has multiple critical gaps — immediately disqualifying Stack 2. Supabase provides fully managed PostgreSQL with native RLS support: free tier at 500MB, Pro at $25/month with 8GB. Data stored in Tokyo for Taiwan latency <100ms.

Authentication: LINE Login and the Taiwan Requirement

LINE penetration in Taiwan exceeds 70%. My LINE community has 466 members who expect to sign in with their LINE accounts. This isn't "nice to have" — it's table stakes. Cloudflare Access: zero support. Supabase Auth: requires custom OIDC configuration (not one-click, but achievable). Clerk: native support for $20/month extra. Stack 4 chose Supabase Auth + custom OIDC, saving cost, with PoC validation as a condition of approval.

Deployment: Why a Single Pipeline Matters

Stack 1 requires 3 independent deployment pipelines; Stack 4 needs just 1. This isn't only a convenience issue — every extra deployment step is another place where AI agent work can break. A single pipeline = fewer cognitive failure points = faster iteration cycles.

5. Why Stack 4 Won with 8.42/10

Stack 4 uses identical infrastructure to Stack 3. Its advantages come entirely from two open-source tools: Turborepo and tRPC.

Turborepo: Monorepo Without the Misery

Turborepo is a Rust-based build system that Vercel acquired in 2021. Its core feature is smart caching (Remote Caching) — if you only changed one UI component, it only rebuilds the affected packages, restoring everything else from cache. This saves 50-90% of build time on incremental changes.

For a one-person team: AI agents may deploy dozens of times per day. Fast build cycles = faster feedback = faster corrections = faster iteration to revenue.

tRPC: The Type System That Lets AI Agents Self-Correct

This is what I believe is Stack 4's most critical advantage — and the hardest to explain. Let me use a supply chain analogy:

A regular REST API is like two departments maintaining separate BOMs (Bills of Materials). If a backend engineer renames a field, the frontend engineer doesn't know — users see a blank screen, discovered 3 days later via customer support.

tRPC is like a single unified BOM for the entire product family. Any inconsistency triggers an immediate red error in the editor — it won't compile, it can't deploy to production. API-related bugs drop by 89-98%.

Here's why this matters specifically for me: the ones writing code are AI agents, not humans. AI agents are excellent at understanding type errors and correcting them immediately. Without a type system, AI has to guess from documentation — and guessing wrong means users pay the price.

6. Coastline Platform Architecture Diagram

7. Monorepo Structure: How the Code is Organized

A monorepo isn't "dumping everything into one folder." It's an organized unified workspace where multiple applications share the same types, components, and configurations.

The critical packages are packages/api and packages/db — all API routes and database models are defined here, then consumed by both apps/web and apps/cockpit. Any type inconsistency throws an error everywhere, immediately.

8. 3-Year Total Cost of Ownership

Tech decisions can't be evaluated on today's costs alone. You need to look at 3-year TCO including infrastructure, development time (AI API costs), and the hidden cost of debugging broken type contracts.

*Stack 2: Lowest infra cost but highest dev cost due to SQLite limitations and custom auth. Disqualified regardless of cost.

Stack 4 and Stack 3 have identical infrastructure costs (both: $0 → $25/month Supabase + $0 → $20/month Vercel). Stack 4 is cheaper because:

• tRPC type safety saves debug time — estimated $700 in AI API costs avoided
• Turborepo caching makes each iteration faster, saving additional AI compute
• Stack 4 is $2,160 cheaper than Stack 1 (42% savings)

ROI context:

• 3 bracelet sales (NT$1,280 × 3 ≈ $120) = 2-4 months of hosting
• 1 workshop enrollment (NT$15,000 ≈ $470) = 8-18 months of hosting
• Traditional dev agency for same deliverable: $50,000-$100,000 (17-33× more expensive)

9. The 5-Judge Panel's Verdicts

This wasn't my opinion alone. I assembled a 5-judge AI panel, each evaluating from a distinct perspective:

10. What This Means for One-Person Teams

I'm not an engineer. But I now have a clear understanding of every layer of this tech stack — not because I suddenly know how to code, but because a rigorous decision process forced me to understand the reason behind every choice.

This reminds me of something my mentor Alex Finn once said: "You don't need to know how to build an airplane. But you need to know why it flies, and under what conditions it crashes."

For a one-person team, tech selection isn't an engineering decision — it's a business decision. The wrong tech stack can trap you in technical debt at exactly the moment you need to move fastest, or force you to rebuild an entire system just as user growth is accelerating.

Stack 4 lets me do the following:

• Day 0 cost: $0 — All free tiers support Phase 1 (0-100 users)
• High AI agent efficiency — Type safety lets AI self-correct; no human code review needed for every line
• No single-engineer dependency — Monorepo + full type safety means any AI agent can take over maintenance
• Taiwan market first — Supabase Tokyo region, latency <100ms; LINE OAuth support
• Open-source protection — Next.js (MIT), Supabase (Apache 2.0), tRPC (MIT). Any vendor failure has an exit path.

One number that stuck with me: a traditional software agency would charge $50,000 - $100,000 to build the same platform. With AI agents and Stack 4, my 3-year total cost is $3,000 — 3% of that figure.

This isn't just about saving money. It's about enabling a one-person team to build infrastructure that was previously only accessible to funded startups. That changes the game entirely.

I'm 51 years old. I don't write code. My family is in Canada and Taipei while I'm building this in Taiwan. And I just made a more rigorous technical decision than most funded startups make. AI didn't replace me — it made me capable of doing this.