What shipped since submission, and two corrections
Work from 2026-09-01 to 2026-10-07, all live at the same URL: 243 commits, model flows from 9 to 21, Gemini 3.7 Flash to 3.8.
Interview prep, first on our "What's next" list:
- What to expect: one click researches a company's interview loop with Google Search grounding; each stage cites a source or is marked inferred.
- A page per round: a brief for its stage, and reported questions, cited only when they match word for word what our model extracted from the cited write-up.
- A six-question typed mock; its debrief quotes back, in amber, claims only you can vouch for.
- Take-home plans from a pasted brief, PDF or link: what the brief asks (quotes checked against it), what people report (cited), a suggested approach (citing nothing).
For applying: a cover letter, cited like any answer, with a letterhead from your profile and a PDF export. It skips web research: you cannot defend a company fact you never read.
A Resume page for the PDF you actually send:
- Diagnose: 14 document checks in code and four yes/no questions per bullet, judged by Gemini. Each row names its source; there is no score.
- Gaps: skills O*NET lists for your role, In Demand or named in your saved postings; one neither evidenced on your resume nor in your profile becomes a question, not a claim.
- Rewrite: a cited replacement bullet, every figure in digits traced to your resume, facts or answers, saved only if you accept.
Also shipped: account self-deletion, a public landing page, an application page that follows the stage, adding or deleting questions, a Copy button, and today's date in date-sensitive prompts, which fixed a tenure miscount.
Two corrections. This product never states a fact about you it cannot cite, and our write-up should meet the same bar; on thinking budgets and temperature it did not. Our measurement notes are private; the code is public.
- Thinking budgets. We wrote that each flow had its own budget: 256 for transcription, 1024 for judgment. The split did nothing: the API treats any budget up to about 3000 as LOW, so every flow, judgment included, ran at LOW, where on 3.8 our real requests spent about zero thought tokens. Each flow now sends a thinking level, MEDIUM or LOW, chosen for the 18 flows we had on 2026-09-14 by a blind evaluation; five of their eight MEDIUMs are hedges on inconclusive evidence.
- Temperature 0. We wrote that Gemini's default temperature made extraction vary and sending 0 made it reproducible. Google's docs had said since 2026-07-21, before we added 0, that temperature, top-p and top-k are deprecated and ignored from Gemini 3.6 Flash on and will be refused by later models, so 0 did nothing on our 3.7 submission; no recorded 3.7 run compared 0 with the default. On 3.8, temperature 0, unset and 2 made no detectable difference over 18 formParse runs, and 0 without a fixed seed gave a short test prompt 5 distinct outputs in 8. Since 2026-10-07 no request carries a sampling parameter; a wire test fails if one returns. Checking output in code, the fix our write-up named, reached the form read only on 2026-10-07: a pasted form's read is now checked for dropped limits in common phrasings and for coming back with under half its questions, and read once more if it falls short.
Live app: https://find-a-job-572064776552.us-west1.run.app Code (MIT): https://github.com/mark24680617/find_a_job Demo video (recorded at submission; what it says at 3:18 about budgets and temperature 0 is corrected above): https://youtu.be/8R0M3HLGvAE
Log in or sign up for Devpost to join the conversation.