How we research our career guides

01 · OverviewWhat you're reading

Every career page on Qoollege is generated by a four-stage pipeline that combines a live web research model, a synthesis model, a structured-JSON transformer, and a text-to-speech model. The pipeline is run by the Qoollege editorial team one career at a time. The same code path produces the article body, the data tables, the source list, and the audio guide, so the surface area you see is consistent across every career.

We're explicit about this because the 2026 web is full of unattributed AI text. If you're going to make life decisions partly on what you read here, you deserve to know what produced it.

02 · PipelineThe four stages, by model

For each career we run the following sequence. Cost and runtime figures reflect our current per-guide averages.

1. Source research — Perplexity sonar-pro. We send a tightly scoped research brief and receive a structured source pack: government statistics (BLS, Census, O*NET), industry surveys, nonprofit and academic reports, and expert commentary, each tagged with a trust category and a URL. This is the only stage that touches the open web; nothing else in the pipeline browses or retrieves new pages.
2. Synthesis — OpenAI gpt-5.4-mini. The source pack is fed to a synthesis prompt that produces a 16-section Markdown report covering snapshot, day-to-day work, pathway, skills, education routes, market, outlook, fit, trade-offs, and student action plans. The model is instructed to ground every claim in the source pack and to flag uncertainty.
3. Transformer — three further gpt-5.4-mini calls. The Markdown report is converted into the structured JSON the website renders: the long-form article object, the summary card fields (salary bands, demand, growth, education), and the high-school and college action plans. Splitting the transform across three calls keeps each JSON schema small and the failure modes localized.
4. Audio — Google Gemini TTS, stitched with ffmpeg. A separate prompt produces a spoken-word transcript, which is then chunked, voiced by Gemini TTS, stitched into a single MP3 with ffmpeg, and uploaded to Vercel Blob. The audio you hear on a career page is the direct output of this stage; no human re-records it.

All intermediate outputs (source packs, Markdown reports, transformer JSON, transcripts) are logged to our database. We can re-render an old guide deterministically from those logs without re-paying for an LLM call.

03 · Human reviewWhat a human checks, and what they don't

We'd rather be honest about this than overclaim. The pipeline is the primary author of every guide on the site. A human does the following:

• Designs and updates the prompts and JSON schemas used at each stage.
• Picks which careers run, and approves the source-pack composition by trust mix.
• Spot-checks newly published guides for obvious factual errors, prompt artifacts, or unsafe advice, and re-runs the pipeline if a guide fails review.
• Investigates any guide a reader flags through the contact form and re-runs or removes it as appropriate.

A human does notrewrite the prose line-by-line, hand-edit the source list, or independently verify every statistic. If we did, we couldn't publish at this breadth. We mitigate this by keeping every claim cited (see below) and by re-running guides on a cadence so figures don't drift far from their underlying sources.

04 · SourcesHow sources are surfaced

Every guide ends with a Research sourcessection that lists each source returned by stage 1, with its publisher, what it covered, and a trust tag (Government, Nonprofit, Industry, Academic, or Expert). Where the source pack carries a URL it is shown as a visit link; where it doesn't (a small number of legacy sources, which we're backfilling) the entry is still listed for attribution.

We weight government statistical sources highest because they're the ones most likely to be checkable and stable. Industry surveys and expert commentary appear where the question (e.g., day-to-day work, hiring signals) is one government data can't answer. The aggregated trust mix is shown in the sidebar of each guide.

05 · LimitationsWhat these guides are not

Qoollege guides are an orientation, not a verdict. A few things to keep in mind:

• Not personalized advice.We don't know your grades, finances, geography, or goals. Salary bands and education routes are national medians and ranges; your local picture will differ.
• U.S.-centric. Salary, demand, and education data are drawn primarily from U.S. sources. International readers should treat the structural sections (what the work is, what skills matter) as more portable than the numbers.
• May be out of date.Each guide is dated at the bottom of its source list. If a labor-market shock or major policy change has happened since that date, the guide hasn't seen it yet.
• AI can be wrong.Even with sourcing, language models occasionally misattribute, paraphrase a statistic incorrectly, or smooth over real disagreement in a field. If a number matters to a decision you're making, click through to the source and verify it there.
• Not financial, legal, medical, or admissions advice. Talk to a human professional for those.

06 · Updates and correctionsHow we keep guides current

Each career page surfaces a "last verified" date pulled from the most recent pipeline run for that career. When a major data source (such as the BLS Occupational Outlook Handbook) publishes an update, affected guides are re-queued and re-run; the new run replaces the old fields and bumps the date. Errors flagged by readers are investigated and, if confirmed, fixed by a re-run rather than a hand-edit, so the fix benefits from the same provenance trail as the original.

Spotted something wrong? Email research@qoollege.comwith the career name and the specific claim. We'll look at it.

07 · VersionThis document

Methodology version 1.0, last updated May 14, 2026. When the pipeline changes in a way that affects what readers see — a new model, a new source-research stage, a change in human review — this page is updated before the change ships, and the version bumps.