Inspiration

All three of us have refreshed an inbox at 2am and found another "we've decided to move forward with other candidates." Or nothing at all. We kept applying out of panic, not strategy, and every tool we used just counted applications. None of them noticed what the search was doing to us.

For those of us on student visas it was heavier. On OPT, every rejection comes with a clock attached. Meanwhile we were handing our résumés and visa status to cloud platforms, and scam "sponsorship" postings were aimed at people exactly like us.

So we built what we wished we'd had: a job search tool that notices when you're stuck and responds like a good friend and mentor would, while keeping your story on your own laptop. Because no one should job hunt alone.

What it does

Unghosted is a job search agent that runs on your laptop. It finds real postings at the companies you choose, scores each one for fit and scam risk, tailors your résumé without inventing anything, applies in a visible browser that stops to ask you when it should, and tracks every reply from your inbox.

Then it does what other tools don't: it adapts to you. A wellbeing engine reads the shape of your search (screens, interviews, rejections, silence) and changes the app as things get harder. Rejections are held for a weekly digest, autopilot slows down, progress numbers step back, and support moves to the top: your own trusted people, real crisis lines for your country and campus, gentle check-ins, guided audio resets, affirmations, and stories of people who've been there.

Everything personal stays on your device. The one optional cloud AI feature only ever sees an anonymous scorecard of numbers, and the app shows you the exact request it sent.

How we built it

Unghosted is a local-first app: a FastAPI + SQLite backend and a React 19 + Vite + Tailwind frontend, running on your laptop. All personal text goes through a local LLM (Ollama, qwen2.5:7b).

The event log is the spine. Every movement of an application (applied, screen, interview, final, offer, rejected, ghosted) is one row in application_events. The apply agent and the inbox write to it, and the wellbeing engine reads from it. One reply email can change what the whole app does next.

  • Search with Tavily + Exa. Queries are built only from your preferences, never from your résumé.
    • Tavily finds the job board of each company you named, limited to the five ATS sites and cached for 30 days. We then list jobs through the boards' public APIs (Greenhouse, Ashby, Lever, Workable, Workday).
    • Exa semantic search adds more postings, restricted to those five ATS domains, so every result is something you can actually apply to.
    • Results are deduplicated by posting, closed postings are dropped, and roles that rule out visa sponsorship are filtered out.
  • Scam scoring. Each posting gets a weighted sum of red flags: free-email recruiter, Telegram/WhatsApp, fee or SSN requests, too-good pay, a board that doesn't match the company, and news red flags found with Tavily web search. $$s = \sum_i w_i\,x_i, \qquad \text{flagged as a likely scam when } s \ge 0.5$$
  • Company briefs. "Check the company" pulls together what the company does, recent news and any red flags, using Tavily and Exa.
  • People to reach out to. Type a role, or leave it empty for recruiters and hiring managers, and Exa's Agent API finds public LinkedIn profiles at that company, so a referral is one message away. Only the company and role are searched, never your data, and each search is capped at $1.
  • Fit score. The local model only extracts a job's required skills. Matching against your profile is deterministic code (synonyms, plurals, implied skills): $$\text{fit} = \frac{|\text{matched}|}{|\text{matched}| + |\text{missing}|}$$ If it can't compute an honest number, it shows none.
  • Tailoring with a grounding guard. Résumé bullets and cover letters are rewritten at three levels. Any number or skill that isn't in your profile is stripped out, and you're told what was removed.
  • Apply agent. A visible Chrome window (Playwright) fills applications on five ATSs, reads every value back, and stops to ask you about custom questions, knockouts, CAPTCHAs and sensitive items.
  • Inbox tracking. A read-only IMAP sync of a dedicated job-search Gmail. Keyword rules run first, with the local model as a fallback, and each classified email becomes an event on the right application.
  • Wellbeing engine. It reads the shape of your search from events, not your mood, and places you in one of four states. Each state changes the app: Rejection Shield holds rejections for a weekly digest, autopilot caps your pace, and support cards move to the top: $$h(\text{state}) = \begin{cases} 0 & \text{steady} \ 7 & \text{stretched or strained} \ 14 & \text{struggling} \end{cases} \text{ days}$$
  • Guided resets with Gradium. Four short audio exercises (box breathing, five senses, "after a no", unclench) are voiced with Gradium text-to-speech in a calm, slowed delivery, with pauses timed to each breath. We recorded them once at build time from our own scripts, so the app plays bundled audio, works offline, and never sends anything to Gradium while you use it.
  • Privacy enforcement. Every outbound request passes an egress allowlist and is logged by host name. Secrets, check-in notes, affirmations and your support circle are Fernet-encrypted at rest, and "Delete everything" wipes it all.

We seeded four personas (Ana, Priya, Sam and Marcus) with realistic histories. Judges can watch the app change live, for example when syncing Priya's inbox moves her from steady to stretched.

Challenges we ran into

  • Automation without lying. It's easy to make an LLM write an impressive résumé, and hard to stop it from inventing a 40% improvement you never made. The grounding guard, and the rule that the model extracts while code decides, came from catching exactly that.
  • Five ATSs, five personalities. Workday needs a separate sign-in for every company, and Lever and Workable sometimes reject automated submits outright. We built the agent to hand the window back to you instead of guessing.
  • Email is messy. "Thank you for interviewing with us, but we've decided not to move forward" mentions an interview and is a rejection. Getting the rule order right decided whether a persona's story made sense.
  • Privacy vs. reality. Job sites need their own scripts and CAPTCHAs, so we couldn't block the apply browser without breaking applications. We record and show every host it loads instead.
  • Caring without being creepy. We had to decide what the app should not show: hiding numbers on hard weeks, never naming your state, and screening crisis language on the device before anything is saved.
  • Four lanes, one codebase. We parallelized by area (search, applying, LLM, wellbeing) and kept a shared CONTRACT.md and fixtures so nobody blocked anyone else.

Accomplishments that we're proud of

  • A full discover → match → tailor → apply → track → adapt loop that runs on a laptop.
  • 575 backend tests, plus a complete offline demo that works with the Wi-Fi off.
  • A Privacy page that shows every connection the app made and the exact request body sent by the one opt-in AI feature.
  • Rejection Shield: the first time we watched rejections arrive as a weekly digest instead of one gut-punch at a time, we wanted it for ourselves.
  • Honest automation: it never fabricates, never solves CAPTCHAs, and never answers a knockout question dishonestly.

What we learned

  • Local models are good at extraction and bad at judgment. Letting the model read and letting code decide made our scores explainable and stable.
  • Privacy is an architecture, not a policy. An allowlist, encryption and a connection log made our privacy claims checkable instead of just stated.
  • Good search is mostly filtering. Deduplication, freshness, closed postings, scams and sponsorship mattered more than the search query itself.
  • What an app leaves out matters as much as what it shows. For someone on their tenth rejection, hiding a metric can be the most helpful feature.

What's next for Unghosted: Finding work without losing yourself

  • Live runs of the apply agent across many more real postings, and better handoffs on Lever and Workable.
  • Career center pilots: universities can add their own campus resources, and the app is privacy-safe to recommend to students.
  • A one-click desktop installer, so nobody needs a terminal to keep their data on their own machine.
  • More boards, plus sharper tools for international students: sponsorship-aware answers and OPT-timeline nudges.

Built With

Share this project:

Updates

Submission history