Designed a company search product and built its frontend

Canonical finds companies from a plain-English description. You describe what you want, it interprets that into criteria you can review, and it returns evidence-backed matches you can inspect and export. It runs on Synaptic's company data. I was the sole designer and product owner, and I built the frontend. The product is live in production ahead of its public launch.

Make it real or remove it.
Every criterion shows its evidence.
No evidence -> the card says so.

Synaptic already had a company-search engine that handled long-tail queries well for its enterprise customers. Exa and Parallel offered public natural-language search that returned weaker results on the same queries.

The brief was to open that capability to a wider, self-serve audience: a plain-English product anyone could use, with public example pages for SEO and credit pricing below the incumbents. It had to reach production fast as an MVP and grow from there. I finalised the fields and filters for the first phase before any UI work, so the product had a defined surface from the start.

Canonical is live in production ahead of its official launch. The product and marketing site I designed and built run in Synaptic's repo, wired to the backend the engineering team built against the contract I prepared.

I owned it from the product definition through to the shipped frontend. The one piece I did not own was the backend: I prepared the API contract from the real UI, and Synaptic's engineering team built against it. I name that boundary on purpose, because it reads as honest scope while keeping the breadth credible.

Most of the product existed on paper before any code. I sketched the first screens in Figma, prototyped the loop in Figma Make, and compared v0, Google Stitch, Lovable, and Replit to settle the flow. The v1 spec went into Notion, every screen got mapped in FigJam, and I built the query interpretation catalogue field by field before opening an editor.

Codex scaffolded the Next.js app: routes, tokens, components, and READMEs. The longer multi-file passes moved to Claude Code once the codebase was real enough for architectural prompts. Every non-trivial feature started as a one-page brief: product context, the policy in plain English, acceptance criteria, and a do-not-invent list. The same brief went to the model, into DECISIONS_LOG.md, and to anyone reviewing the code later.

The frontend is real Next.js running in Synaptic's production repo. A few changes show what that meant in practice.

The tools helped most on cross-file refactors, typed content systems, and audits. They helped least on anything that needed taste or a first-principles call. Left alone they reached for a flat table palette, generic visuals, and find-and-replace copy edits, so I supplied the colour maths, a ban list on the visuals, and sentence-level rewrites instead.

I owned the product. Codex and Claude Code executed against briefs I had already written. Several of the biggest calls were removals: a clarification panel that did not move the search, a redundant marketing section, an unaudited compliance badge, and copy rewritten rather than character-swapped. Each removal made the product more honest.