oppachikin.

01

The problem

My sister runs a fried chicken shop in Minsk, Belarus. Unlike China's "everything-on-Meituan" delivery ecosystem, Minsk has weak local delivery platform coverage — most restaurant orders still come through walk-ins, phone calls, and the shop's own website. A website here isn't a bonus channel — it's the primary acquisition channel.

She hired me, paid me a development fee, and gave me my first real freelance gig.

02

My angle

I was a financial analyst at SF Tech with no formal frontend/backend experience. My bet going in: using Cursor + Claude Code, I could translate my understanding of business and users into working code. AI handles syntax, library choices, best practices; I handle product judgment, debugging intuition, and the feedback loop with a real customer (my sister).

03

Architecture & key decisions

A decision I'm proud of: loyalty points are stored as an append-only transaction log, not as a mutable balance.

Direct instinct from my finance background — anything money-shaped needs an audit trail. When something goes wrong, you can replay every change, not just stare at a wrong number.

04

The hardest (and proudest) parts

None of this was built up front — I dedicated a full week near launch to a security audit and patched these one by one. The commit log shows a week of consecutive security: prefixed commits. This is the part of the project I'm most proud of.

05

AI collaboration workflow

Day-to-day coding in Cursor. Architecture-level decisions: discussed in Claude Code chat, weighing trade-offs, then generating code. Complex integrations (e.g. Alfa Bank): fed the gateway docs into Claude Code, had it produce a doc-grounded integration plan, then debugged together.

[Specific vibe-coding moment — TBD: one feature I had zero intuition for, that AI helped me decompose and ship.]

The biggest lesson: AI doesn't make product decisions for me. It makes the gap between "an idea I can articulate" and "a working implementation" 10x narrower. The bottleneck becomes how clearly I can think. That muscle came from finance, and it transferred cleanly.

06

Current status

• ✓Engineering: done
• ✓Security hardening: done
• ⏳Launch: awaiting SMS.by sender name approval (local compliance, not engineering)

07

What these 5 weeks taught me

1. AI lowers the "can I do this" bar but raises the "have I thought it through" bar. My finance background turned into an asset — I'm trained to define problems before acting.
2. Production-grade ≠ feature-complete. Features are the start; idempotency, validation, webhook signing, rate limiting, secret management — a long tail.
3. A paying customer is the best product teacher. My sister's reactions to the UI taught me more than any tutorial.
4. What's next: bring this "AI collaboration + business sense" combo into companies trying to use agents to improve internal workflows.