AvsB.Experiments,end to end.

AvsB is a full-stack A/B testing and feature-flag platform — the kind of product companies usually license from Optimizely, LaunchDarkly, GrowthBook, or Statsig.

It covers the full loop: build experiments in a visual or code editor, ship them through a zero-dependency snippet, ingest events at the edge, and read results back across three different statistical engines — frequentist, Bayesian, and sequential.

Sole developer. Started over the 2024 Christmas holiday and have been at it for 1.5+ years since. A couple of earlier versions were scrapped along the way — early architectural decisions that didn't hold up at the scale the product was growing into. The current iteration spent its full first year on foundation work, on the strength of what those false starts taught me.

Built with Claude Code as a daily co-engineer through the 1.5+ year build. At this scale, the skill isn't typing prompts — it's running AI well enough to actually ship: specs before code, tasks scoped tight enough that the agent can deliver, multi-file refactors and schema migrations delegated, critical paths hand-reviewed. A 410k-line solo platform is finishable when the workflow around the AI is itself a craft.

Experimentation platforms have to be three things at once: a comfortable surface for the people who run tests, a fast and cheap analytics pipeline, and a statistically defensible analysis layer.

Most commercial tools nail one and compromise on the others. AvsB is an attempt to take all three seriously in a single, coherent codebase — without giving up the developer ergonomics of code-first workflows.

A coherent platform across six surfaces, sharing one statistics engine and one analytics backbone.

• Experiment builder with a real visual editor
• Feature flags with per-environment targeting and A/B-test rules
• Three statistical engines running side by side, with variance reduction and always-valid analysis
• Edge ingestion across five Cloudflare Workers, in front of a ClickHouse store
• A full SDK family — browser snippet, JS / Node / React, CLI, browser extension
• The supporting layer most experimentation tools take for granted: orgs, roles, audit log, billing

Chosen for fast iteration and a hosting model that stays cheap as event volume grows.

• Next.js 16 (App Router), React 19, TypeScript (strict)
• Prisma + PostgreSQL · ClickHouse for analytics
• Cloudflare Workers + Durable Objects · R2 for storage
• Redux Toolkit · NextAuth · Zod · SCSS Modules
• Vitest · Playwright · MSW · Testcontainers

• ~410,000 lines of TypeScript / SCSS
• 250 API routes · 71 Prisma models
• 867 test files · ~8,000 test cases
• 5 Cloudflare Workers · 4 published SDK packages

Solo build, 1.5+ years in. Currently pre-launch — no external users yet — while the public API and pricing stabilise.