Selected Work

AI writing tool (fanfic generator)

Problem

Ship a real paid AI product: free usage that can't be farmed, model spend that can't run away, and content rules strict enough for payment providers.

Delivery

Streaming generation, a free tier plus credit-paid model tier, story continuation, two-layer content safety, cost tracked from real token usage, and Waffo / PayPal / Stripe payments — on my own Novalith template, live in three days with Claude Code.

Evidence & boundary

Live at fanficwizard.com; all three payment providers verified end to end in test mode. Too new for traffic or revenue data.

Personal brand and overseas delivery portfolio

Problem

Turn engineering experience, service offers, proof of work, and project intake into one credible site that can keep evolving.

Delivery

Independently shaped the positioning, bilingual content structure, Next.js implementation, SEO, responsive QA, and launch maintenance.

Evidence & boundary

The website itself is the accessible delivery artifact. Lighthouse 100 across the board on desktop; 96 / 100 / 100 / 100 on mobile.

Subscription + credits SaaS operations console

Problem

A product that sells subscriptions plus usage credits—the model most AI products use—can keep everything in PostgreSQL and still lack shared metrics, safe credit changes, a risk investigation queue and a trustworthy record of who did what.

Delivery

One internal console for subscriptions, a credit wallet with an immutable ledger, orders with idempotent mock payments, risk cases, five-role RBAC and an append-only audit log—every write behind the same session, permission and rate-limit guard.

Evidence & boundary

Capability case on synthetic data. 65 tests, 6 E2E tests and 17 security and risk acceptance checks passing.

AI feature / RAG / agent / MCP

Problem

Many AI support bots are a chat box on a prompt: no visible evidence for a conclusion, and nothing that stops an action they shouldn't take.

Delivery

Cited RAG, an independent MCP tool server and a bounded agent state machine; high-risk actions wait for a signed, one-time human approval. Model vendors are swappable.

Evidence & boundary

Capability case for a fictional store. 98 tests and 12 evaluation cases passing; verified live with DeepSeek and Qwen embeddings.