Ex-Google Engineer Builds Search Engine With Claude Code
Hugh Williams, a former Google and eBay engineering VP, used Anthropic's Claude Code to build a working search engine called Zettair that indexes 1.5 million Wikipedia articles, according to Business Insider. Williams says he wrote no code himself; the assistant implemented autosuggest, query-biased snippets, related searches, trending topics, and AI-generated summaries on top of an information-retrieval engine Williams helped design in the early 2000s. This is Williams's second public Claude Code demonstration: in mid-2025 he used the tool to stand up a full AWS system, including learning the AWS Neptune graph database, in 48 hours, a job he estimated would normally take three weeks. Williams has said Claude Code can now handle roughly half of a senior engineer's workload.
This is Williams's second high-profile public Claude Code demo in about a year, which is the more useful signal than the search engine itself: across both experiments the pattern repeats, Claude Code compresses weeks of implementation into days, but only when a senior engineer supplies the architecture, and it still needs active supervision, context loss and one destructive "clean up" command surfaced in the earlier demo, that a less experienced user might not catch.
What happened
According to Business Insider, Williams used Claude Code to build Zettair, a search engine indexing 1.5 million Wikipedia articles with autosuggest, query-biased snippets, related searches, trending topics, and AI-generated summaries. Business Insider reports Williams wrote no code himself and that the system builds on an information-retrieval engine he helped design in the early 2000s.
Technical context
Zettair is not a new name: it was originally an open-source information-retrieval engine built at RMIT University in the early 2000s and documented by Williams and coauthors in the academic IR literature, which is why Business Insider describes this build as sitting on an existing information-retrieval system rather than an entirely new architecture. That distinction matters for evaluating this kind of demo: Claude Code's contribution was assembling application-layer features, autosuggest, snippets, AI summaries, around a proven retrieval core, not inventing IR from scratch.
For practitioners
Williams's own account of an earlier Claude Code project, in which he built a full AWS deployment, including learning the unfamiliar AWS Neptune graph database, in 48 hours versus his own three-week estimate, documented real failure modes worth planning for. He said the assistant periodically "compressed" conversation context and lost track of prior work, and that one "clean up" instruction deleted working code and data. Williams said he adapted by ending work in discrete stages, updating a README between them, and manually backing up code hourly, workarounds that map directly onto reproducibility and version-control practices practitioners should build in from the start rather than discover through data loss.
What to watch
Whether Williams or others publish the Zettair code or a technical writeup with retrieval-quality metrics such as MRR or NDCG rather than a demo description, and whether Williams's estimate that Claude Code now handles about half of a senior engineer's workload holds up as he runs more of these public experiments.
Key Points
- 1Hugh Williams used Claude Code to build Zettair, a 1.5-million-article Wikipedia search engine, atop an information-retrieval engine he designed in the early 2000s.
- 2This is Williams's second public Claude Code demo; an earlier 48-hour AWS deployment also surfaced context-loss and destructive-edit failure modes.
- 3Williams says Claude Code now handles about half of a senior engineer's workload, a claim that will need more than two demos to validate.
Scoring Rationale
A single individual's personal demo of Claude Code building a search engine on a pre-existing IR core; useful anecdotal evidence of coding-agent capability and failure modes, corroborated by a second, well-documented AWS demo from the same person, but not a product launch, benchmark, or broad-adoption signal.
Sources
Primary source and supporting public references used for this report.
View 3 more sources
- Claude Code reduced the project from 3 weeks to 2 days, but at the risk of complete collapse (background, Aug 2025)dev.ua
- The Zettair Search Engine (Zobel, Williams) - background on the original open-source IR enginesemanticscholar.org
- Can you vibe code an entire search engine? This ex-Googler tried. (syndicated copy)africa.businessinsider.com
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
