Intelmaxxing

LinkedIn shows you yesterday's listings. IntelMaxxing shows you next month's hiring intent five AI agents reading eleven live sources to predict the job before it's posted.

Comment

Inspiration

Every career tool at this hackathon was a wrapper on LinkedIn. Another feed of jobs that were already posted, already flooded, already yours-if-the-recruiter-even-opens-your-résumé.

The real edge isn't another listing it's seeing hiring intent before the listing exists. A Series A closes Tuesday. The founder tweets about doubling the team Wednesday. The job post goes live six weeks
later. By then, you're applicant number 400.

I wanted to land in the founder's inbox on Wednesday.

What it does

IntelMaxxing is a career intelligence pipeline. You describe a mission in one sentence. Five AI agents deploy in parallel across eleven live sources. A full dossier lands in about twenty seconds.

  • FOXHOUND — funding rounds, YC batch moves, Show HN launches
  • WIRETAP — Hacker News "Who is Hiring" comments pulled live
  • GHOSTNET — GitHub repos worth contributing to (commit velocity, good-first-issues)
  • DOSSIER — per-company case files with founder emails, Reddit chatter, and a draftable cold email
  • PROFILER — cross-references everything and narrates a noir voice briefing

Type /case and a manila folder peels open with every public engineer ranked by commits, every guessed email pattern, and a Reddit red-flag sweep — ghost interviews, rescinded offers, layoff chatter surfaced before you send the email.

How I built it

  • Next.js 16 + React 19 + Tailwind v4 on Vercel Fluid Compute
    • Gemma 4 31B via Novita for orchestration · ElevenLabs for voice
    • Five agents fan out behind Promise.all, each with its own deadline
    • Every source is bounded — one slow API never stalls the mission
    • 90-second LRU cache so identical prompts stay stable across re-runs
    • Loading isn't a spinner — it's an animated signal graph with rolling transcriptions and reassurance copy that fades in at 12s / 20s / 30s

Challenges I ran into

The retry trap. My LLM wrapper retried once on timeout. Seemed helpful — until a single retry burned fifty seconds inside a 28-second agent deadline and returned empty. Lesson: never retry when the caller
owns a fallback. Retries stack, deadlines don't.

Determinism. Same prompt, ten seconds apart, gave different results. Judges would read that as broken. Fix was three layers: temperature 0 on the classifier, an LRU cache on the payload, and normalized cache keys. Dead time. The 20-second wait is where every demo dies. I spent a disproportionate amount of polish budget there. Per-agent elapsed counters, intel counts that ease from 0 to target, a progress bar that never hits 100 until the data actually lands. Turned dead time into content.

Accomplishments that I am proud of

Solo build, 24 hours. Five agents, eleven sources, animated case-file reveal, voice briefing all shipping on one URL: intelmaxxing.tech.

  • Twenty seconds, not ninety. The pipeline used to take over a minute. Aggressive parallelism, per-source deadlines, and a hard 12-second LLM ceiling got it down to a demo-reliable ~20s.
  • Zero dead whitespace. Every 20-second wait is content — agent transmissions, network-graph hops, intel counts animating, reassurance copy fading in on the 12/20/30-second marks.
  • The red-flag layer. Saving users from weeks of their life. If a company has ghosted three candidates this quarter, we surface it before the cold email goes out. That's the feature that feels like it
    matters.
  • The /case command. A full operational dossier on any company the judges name engineers, founders, emails, Reddit chatter, cold email in under two seconds, with a 3D manila-folder reveal.
  • Graceful degradation is a feature. Any single source can fail and the rest still ship. That's the thing that makes this demo-safe in a convention-center Wi-Fi.

What I learned

  • Per-source deadlines beat global timeouts. A global timeout lets the slowest source pace everyone. Per-source deadlines mean the fast ones ship on time while the slow one drops off quietly.
    • Deterministic fallbacks make ambitious architecture safe. Every Gemma call has a rule-based fallback, so five parallel LLM calls don't turn into five parallel failure modes.
    • Never retry inside an LLM wrapper. Retries stack on top of outer deadlines and turn a 12-second ceiling into a 50-second hang.
    • Cache the repeatable. A 90-second LRU on identical prompts is cheap insurance against judge-visible flicker.
    • The loading state is the product. Products that feel fast reframe the wait as content. Not a spinner. Not a progress bar. Content.

What's next for Intelmaxxing

  • Chrome extension — right-click any company, open the dossier in a side panel
  • Slack bot — ping me when a target drops a Show HN or closes a round
  • Free tier forever — career intel shouldn't live behind a recruiter's retainer

Built With

Share this project:

Updates