Qoollege Career Research

Problem

Career information students rely on is scattered, shallow, or out of date, and producing trustworthy, well-structured, cited guides by hand doesn't scale. A content site also has to be genuinely fast and discoverable — most are neither — or the research never reaches the students searching for it.

Approach

We built a two-part system in a Turborepo monorepo over one Neon Postgres database. A Python research pipeline handles one career at a time: Perplexity sonar-pro assembles a sourced research pack, OpenAI drafts a 16-section structured report, a three-way transformer splits it into article / card / actions JSON, and Gemini TTS plus ffmpeg render a narrated audio version uploaded to Vercel Blob — with every run and its structured fields written to Postgres (careers + pipeline_runs). A Next.js 16 server-rendered viewer reads from Postgres and ships SEO-first career pages with structured data, a persistent audio player, compare views, and a subscriber digest. The pipeline runs offline so no AI cost or latency ever sits on the request path.

Outcome

Live at careers.qoollege.com. The viewer scores Lighthouse 100 performance / 100 SEO / 100 best practices on desktop (92 / 94 / 100 on mobile) with accessibility lifted to 94 after a targeted pass, and Core Web Vitals comfortably green — LCP 0.6s, CLS 0, TBT well under budget. Each career renders as a server-rendered, schema-marked page with an audio 'podcast' version, so the AI-generated research is both fast for students and legible to search and answer engines.