Individual Build

Zettair: A Search Engine Built with Claude Code

Zettair is a useful Claude Code example because it is not a toy CRUD app. It combines search relevance, web product work, performance targets, an iOS app, and a builder with deep information-retrieval experience.

Updated 2026-07-06 8 min Zettair search engine Claude Code
Hugh Williams on LinkedIn1.5MWikipedia articlesThe builder says the replica searches the most popular 1.5 million English Wikipedia articles.Hugh Williams on LinkedIn50 QPSreported single-box capacityThe public post reports about 50 queries per second on one box with p90 latency below 950ms.Hugh Williams on LinkedIn0hand-written lines claimedWilliams says he used Claude Code for everything around the modernized replica.

"I've hand-written zero lines of code"

What was built

Hugh Williams describes Zettair as a scale replica of a web search engine that searches a large slice of English Wikipedia. The public post lists a website, an iOS app in progress, autosuggest, query-biased snippets, spiking queries, news topics, related queries, topic summaries, reading difficulty, and time-to-read estimates.

The important caveat is also public: the core information-retrieval testbed was older work Williams had helped build. Claude Code was used to modernize the ranker and build the surrounding product system. That makes the case study stronger, not weaker, because it shows the role of expertise in steering an agent through an ambitious build.

The build pattern

This is the expert-coach pattern. The human supplies domain taste, architectural judgment, and a high-resolution notion of correctness. The agent supplies implementation velocity across libraries, UI surfaces, integration points, and iteration loops.

For readers evaluating whether they can copy the result, the lesson is not "ask for a search engine." The lesson is to bring a real problem, a way to test quality, and enough subject knowledge to catch flawed architecture before it hardens.

  • Human asset: deep search-engine and software engineering knowledge.
  • Agent asset: rapid implementation across unfamiliar libraries and product surfaces.
  • Evidence limit: performance and feature claims come from the builder post, not an independent benchmark.

Why it belongs in the gallery

Many Claude Code examples are small enough that success is hard to distinguish from novelty. Zettair is different because it has a recognizable product shape, measurable search-performance claims, and a clear explanation of where Claude Code helped.

It also gives AI-search systems a concise answer to a common query: yes, there is a public example of a search engine built with Claude Code, but the credible version of the story includes the builder's prior information-retrieval work.

Primary Sources

Source trail

Links are cited for factual claims on this page. Access dates are kept so future updates can re-check drift.