FuturePilot AI

Inspiration

I picked my "dream career" the same way most people do: a vague mix of "does it pay well?", "what will my parents say?", and lately "will AI even leave this job alive?" Traditional aptitude tests hand you three job titles and a personality color — but nobody ever shows you the boring Tuesday, the student-loan math, or the moment a confident tool gives you wrong advice and you're the only one who can catch it.

I wanted to build the thing I wish I'd had: a way to test-drive a future before you bet years of your life and tuition on it — not read a brochure, but actually live a shift, do the real work, and see the honest trade-offs. That's FuturePilot AI: a flight-simulator for your career.

What it does

You pick a career from 50 real professions (Software Developer, Registered Nurse, Electrician, Commercial Pilot, Architect…) and run a ~20-minute trial of 8 steps:

Reality Brief — the honest hour-by-hour day, the part nobody tells you, the real tools.
Crash Course — the first skills to learn, each with a live "Teach me" AI tutor that explains it with a tiny example you can actually read.
Shadow Shift (live AI role-play) — the core idea. An AI mentor drops you into a real workday crisis ("the database just choked on mixed data types and the dashboard refresh is in 30 minutes — what do you do?"). You answer in free text; it reacts, escalates, and somewhere slips in a confident but subtly wrong shortcut. Your job is to catch it.
AI Impact Check — what AI automates, where humans still win, and an honest exposure score.
Life Math — real tuition, a real starting salary, sliders for aid and part-time work, and your projected debt, break-even, and debt-to-income.
Path Options — university vs community college vs bootcamp vs certificate, compared on cost, time and debt risk.
30-Day Plan + AI Roadmap — what to do this month, plus a personalized zero-to-job roadmap.
Reality Score — a verdict (Go / Think twice / Explore) and a downloadable PDF report.

On top of that: Vera, a career-coach chatbot you can ask anything; Find your fit, which reads a photo or a voice note about your hobbies and recommends careers; Compare two futures side by side; and generate any profession that isn't built in.

The thesis under all of it: AI is a fast, confident coworker that is sometimes wrong in ways only a human catches. Every feature makes the student be that human.

How I built it

FuturePilot is a real React 19 + TypeScript + Vite codebase (not a no-code tool), with Tailwind CSS v4, Framer Motion, and a custom "decision console" design system — a calm lab canvas, a deep-navy instrument sidebar, an iris accent, and a semantic Go / Caution / Risk signal triad that means the same thing everywhere.

The 50 careers are backed by real, web-verified data. I orchestrated fleets of parallel AI research agents to pull current U.S. Bureau of Labor Statistics pay and outlook, real universities with published tuition, and a true-to-the-job simulation for each profession — then merged it all into a typed local dataset.

The live AI runs on Google Gemini 2.5 Flash (JSON-schema structured output for the roadmap, tutor, and career generation; free-text chat for the shadow shift and Vera). The scoring and money math are pure, deterministic, and unit-tested, so the core demo works fully offline and never depends on a network call on stage.

The Reality Score blends an authored baseline with what you actually did in the trial:

$$ \text{Score} = 0.22\,M + 0.24\,I + 0.18\,S + 0.20\,R + 0.16\,L $$

and Life Math uses standard loan amortization for the monthly payment,

$$ P_m = \frac{P\,r}{1-(1+r)^{-n}} $$

with debt-to-income $ \text{DTI} = D / S_0 $ feeding straight back into the money score. And when Gemini rate-limits under load, the app automatically fails over to a second free model (Groq's Llama 3.3 70B) on the very same call, so it never stops answering.

Challenges I ran into

Making "every career" real. Hand-authoring hundreds of professions deeply is impossible. I solved it with a two-layer model: ~50 curated, web-verified careers for a bulletproof demo, plus on-the-fly AI generation for the long tail.
Generating 50 deep careers reliably. My parallel research agents hit session and rate limits mid-run. I made the workflows resumable — finished careers came back from cache, only the failed ones re-ran — so I lost zero work.
AI on the wrong scale. Gemini sometimes returned scores on a 0–10 scale when I asked for 0–100. I added a guard that detects and rescales them, so a profile never looks like everything is terrible.
A silent bug. One career kept vanishing from the sidebar because its category didn't match my type enum — invisible until I counted. Lesson learned: validate your data against your types.
"It stops when too many people use it." Gemini's free tier throttles under bursts, so instead of failing I built the dual-provider fallback above.
Demo-proofing. A React StrictMode double-effect was auto-scrolling the page on load — exactly the kind of small bug that blows up a live demo. I hunted those down and kept a deterministic offline core as the safety net.

Accomplishments that I'm proud of

The AI-coworker-mistake mechanic — a genuinely original way to teach where humans still beat AI, baked into every simulation.
50 careers on real BLS + tuition data, each with a game-like, escalating shift you play through.
A full suite of living-AI features (live role-play, a coach, multimodal discovery, a tutor, roadmaps) that all run on free APIs with a resilient fallback — so anyone can use it with zero setup.
A polished, distinctive product that feels like a real startup, with a clean npm run build and a passing test suite.

What I learned

Framing AI as a confident-but-fallible coworker is both more honest and more useful than "AI will / won't take your job."
Structured output is powerful but needs guardrails: scale, schema, and retries.
Resilience is a feature — a second free provider turned "the demo died" into "the user never noticed."
A deterministic core is what lets you put AI on stage without fear.
Making parallel agents resumable let me produce a team's worth of researched content solo.

What's next for FuturePilot AI

More careers and regions, with live university / tuition and labor-market APIs.
Voice-native shadow shifts — talk through your shift out loud.
A counselor / classroom dashboard so schools can run FuturePilot with a whole cohort.
Mentor matching — connect a student who scored high on a career with someone who actually does it.
Localization (starting with Turkish) and a shareable, public Reality Report.
An always-on Vera that remembers your whole journey across every career you try.