HOW IT WORKS
From job boards to a grounded draft.
A daily data pipeline feeds the search. A guarded AI demo tailors an application to any posting for a fictional candidate, and every claim it makes points back to a fact.
Live numbers
Open postings175.5KSearchable right now
Companies5,007
New since yesterday5,940
Data last rebuiltOct 1002:00 UTC, daily
Demo runs19Last 30 days
Served from cache63%Instant and free
Cost per fresh run1.7¢Fit check plus draft
Typical fresh run16 sMedian, streamed live
Architecture
Every day: the job data
- Company job boardsGreenhouse, Lever and other public boards
- ScraperA separate pipeline collects postings and descriptions
- dbt models in PostgresCleaned and classified by role family and level
- Read-only readerThis app can only SELECT two marts, nothing else
- SearchFilters, multi-city grouping and stored descriptions
Every demo click: the tailoring
- Cache checkA posting someone already ran replays instantly
- GuardrailsPer-visitor and global limits, and a daily dollar cap reserved before sending
- Fit checkIs Alex Sample a fit? Strict JSON, citing fact IDs
- Fit gateNot a fit stops here: no letter is written or paid for
- DraftRésumé bullets and a short cover letter, each citing facts
- ValidatorDrops or flags anything uncited or claiming a tool Alex never used
- Claim checkA second, cheap call tests each sentence against the facts it cites; anything unsupported is highlighted
- Stream and cacheProgress streams to the page; the result is saved for next time
Why you can check every claim
- Alex Sample has 22 numbered facts. The model can only cite IDs from that list, and the app re-checks every one.
- Requirements are quoted from the posting; paraphrases are labeled.
- A list of tools Alex has never used is checked in every sentence. "I haven't used Kafka" is fine; "Ran Kafka" is not.
- A second model call checks every sentence against the facts it cites and highlights anything they don't back up.
- Unrelated roles get an honest "not a fit" instead of a stretched letter.
Why it costs cents, not dollars
- Results are cached per posting and prompt version, so a repeat click costs nothing.
- The fit gate skips the draft call for roles that don't match.
- Every request reserves its worst-case cost against a daily cap before it is sent, then settles to the real cost.
- Newest postings are pre-generated overnight, but only after real visitors have used the demo.
The private workflow
Beta members use the same ideas on their own background: confirm facts from a résumé, pick a posting (its description is saved with the workspace), read the employer's questions, then draft, review and export. Writing samples teach tone, never qualifications. The applicant always submits on the employer's site.