Unbias JD
Web app that scans job adverts for biased or exclusionary language and suggests neutral rewrites. Combines a lexicon scan with a contextual LLM pass via Vercel AI Gateway. English and Japanese, stateless.
Unbias JD reads a job advert and reports where the wording quietly excludes people — gendered phrasing, age signals, ableist defaults, nativism, culture-fit shorthand, elitist credentialism — then offers a neutral rewrite. English and Japanese.
Detection runs in two layers rather than one. A lexicon scan catches known phrases deterministically, fast and free; a contextual LLM pass through the Vercel AI Gateway then judges what a word list cannot — whether a phrase is actually exclusionary in this sentence. Splitting them matters because each covers the other's blind spot: the lexicon never hallucinates, the model never misses a novel phrasing.
The two layers also degrade independently. Without an API key the app still runs the lexicon scan and surfaces a banner explaining that the contextual pass is unavailable, instead of failing shut. URL fetching goes through Mozilla Readability with SSRF protection, and the AI-backed routes sit behind Vercel BotID so the expensive path is not free to script against.
There is no database. Pasted text is sent to the gateway for the contextual pass and is never persisted by the app — a deliberate constraint, because the input is often an unpublished job advert and asking someone to upload that to a stranger's server is a bigger ask than the tool is worth.
Stack
