In your Words

Job applications in your own words. Nemotron drafts your screening answers, then checks every sentence against stories you said aloud. A line you never said is cut or asked about.

Comment

Inspiration

A strong senior job application takes me 30 to 60 minutes. Plain AI drafts are fast, but they make up numbers and scope. Then I spend the saved time finding what is not true. The slow part is not writing. It is trust.

What it does

You say your career stories once, by voice. Each line gets a number. For each role, In Your Words reads the posting and the company, drafts your screening answers and a short note, and checks every sentence against your stories.

  • Each kept sentence shows the story lines it came from.
  • A line your stories cannot back is cut, and you see why.
  • When a claim has no proof, the tool asks you one question. Your spoken answer rewrites the line, and it is saved as a new story line.
  • A defend sheet lists every claim next to what you actually said, so you are ready for the interview.

How we built it

  • Nemotron 3 Ultra on Nebius Token Factory drafts. Each sentence states one kind of fact and names the story lines it rests on. - Nemotron 3 Super on Token Factory judges every sentence against those lines and the lines near them. It returns kept, cut or flagged, with a reason.
  • Tavily extracts the job posting and searches the company at runtime.
  • Plain code checks every number in a kept sentence against its source. A model never gets the last word on a number.
  • Next.js 16 on Vercel, Neon for a daily live-run counter, PostHog for anonymous page events. No agent framework, no vector store.

Challenges we ran into

  • With thinking on, Ultra spent its whole 12k reasoning budget before any answer. I run it with thinking off.
  • Quote matching alone let a false claim through in my first test. So the judge checks every sentence, and it sees nearby story lines too.
  • Checking takes time. On my real run, reading the posting took 5.9 s, drafting 5.1 s, and checking every sentence 35.7 s. The check is the slow step, and it is the step that matters.

Accomplishments that we're proud of

  • Every sentence in the output is backed, labelled as inference, or cut with a reason.
  • A full package costs about $0.02 to $0.03 of Token Factory credit.
  • I sent a real application with it on 30 Oct 2026. 18 sentences were checked: 17 kept, 1 cut. I answered the open questions by voice, and those lines were rewritten from my words. From posting to sent took 8 min 51 s. The video shows that run, with my own text blurred.

What we learned

  • A judge model is only as good as the evidence you hand it. Giving it nearby lines mattered more than prompt tuning.
    • Prompt tuning on the judge showed no measured gain on real lines. I stopped and kept the simpler prompt.
    • I measured honestly. On my first real role, a plain AI draft was faster to produce (224 s) than a full checked package. The checked package is the one I could defend line by line.

What's next for In your Words

  • A claim ledger: one list of every fact and number you have stated, across every application.
  • Same story everywhere: a check that you tell each story the same way in every application.
  • Say it back: speak your answer before the interview, and any claim you cannot say without reading gets flagged.

Built With

Share this project:

Updates

Submission history