Inspiration
What it does## Inspiration
Every job post costs an hour to do properly: read the full description, check fit honestly, tailor a one-page resume without inventing anything, find a real person to ask. I am an international MSc student in Singapore with 12 hours of class, a 16-hour part-time job and research every week. So most posts I find die in my camera roll.
What if that hour took one minute, and every claim on the resume was still true?
My own deployment has already collected 265 postings from 31 sources, scored 104, and helped me pick 14 applications.
What it does
One chat. Drop anything in. A screenshot, a careers link, pasted text, or a plain question. A Nemotron 3 Nano router reads intent in about 1 second, runs the best expert, and offers the next likely steps as one-tap buttons. No slash commands, no forms.
Three experts, each with its own context:
- Job Fetcher. Nemotron 3.5 Super VL reads the screenshot. Tavily Extract pulls the full description from a link. Nemotron 3 Ultra scores the job against my profile with hard gates: eligibility, scope, timing, hours, exclusions.
- Resume. Ultra drafts a one-page resume from my master resume. Then code checks every number against the master, the bullet rules and the page length. Ultra gets one chance to fix what failed. The model writes, code verifies.
- Outreach. Tavily searches public pages, LinkedIn always excluded. Nano proposes a person, and code accepts them only if the quoted evidence is really on that page. Ultra drafts the email and a 200-character note; Nano reviews it. No verified person? It says so instead of inventing one.
Always on. Every morning at 05:00 a cron job collects new postings, Nano triages them and Ultra re-checks the best ones.
You stay in control. Drafts only. Every outward action needs your tap. A /memory page shows exactly what each expert can and cannot see.
How we built it
Next.js 15 on Vercel, Neon Postgres, Nebius Token Factory for every model call, Tavily Search and Extract.
Right model for each job, picked by role: | Role | Model | Used for | |---|---|---| | fast | Nemotron 3 Nano | routing, triage, self-checks | | reason | Nemotron 3 Ultra | scoring, resume, drafting | | vision | Nemotron 3.5 Super VL | screenshots | | fallback | Nemotron 3.5 Lightning | errors and timeouts |
One client records model, latency, tokens and USD for every call and enforces daily and total spend caps. At 60% of the cap, Ultra steps drop to Nano with a visible banner, never an error. Tests check that the router never sees the profile or the master resume. 188 tests run in CI with fakes, so building it spent zero credits.
Measured on the live deployment
| Flow | Models | Time | Cost |
|---|---|---|---|
| Screenshot to scored lead | Super VL + Ultra | 21 s | $0.008 |
| Job link to scored lead | Tavily + Nano + Ultra | 24 s | $0.006 |
| Tailored resume PDF | Ultra + gates + 1 revision | 25 to 34 s | $0.037 |
| Referral draft | Tavily + Nano + Ultra + Nano | 14 to 25 s | $0.005 |
| Daily run | Nano triage + Ultra top 2 | $0.015 |
The whole live verification cost about $0.15.
Challenges
- Keeping a resume honest. The gates report a failed check plainly ("79% of bullets quantified, need 85%") instead of hiding it.
- Structured output. Strict json_schema on Nano returned malformed JSON for our router schema. JSON mode with the schema in the prompt was clean every time.
- Hidden reasoning. Nano, Super VL and Lightning think by default, which used up small max_tokens budgets and returned empty content. Turning thinking off per call fixed both latency and cost.
- Never touching LinkedIn. Every Tavily call excludes linkedin.com. A LinkedIn link gets "send a screenshot instead".
What we learned
Small models route, big models reason, code verifies. That split made the product cheap enough to run every day and honest enough to trust.
What's next
- Accounts for everyone: upload your master resume first, and every expert works from your data in its own workspace.
- Telegram as a second channel on the same orchestrator.
- Memory edits from chat ("stop showing me agent roles").
- PDF and docx inputs.
How we built it
Challenges we ran into
Accomplishments that we're proud of
What we learned
What's next for AITriceps
Built With
- nebius-token-factory
- neon
- next.js
- node.js
- nvidia-nemotron
- pdf-lib
- postgresql
- react
- tavily
- typescript
- vercel
Log in or sign up for Devpost to join the conversation.