I built a model that rates careers like stocks — then caught it lying, and fixed it.
by Ashley Libasci — Math + CS @ UIUC · GitHub · LinkedIn
CareerStar gives every U.S. career a single 0–100 score for how viable it is in an economy AI is reshaping — the payoff (growth and pay), marked down by the risk that AI eats the work, the same way an investor marks down a promising-but-risky stock. It’s a solo summer project, built with an AI-native workflow and deployed on AWS. This page is the story of how it works and why I trust it.
What is this page & how do I use it?
This is the story behind the product — how the model was designed, the bug I caught in my own methodology (and chose to disclose rather than hide), and the evidence that the scores hold up. It’s written for anyone deciding whether to trust the site, including recruiters kicking the tires.
How to use it: read top to bottom like an essay — or jump straight to the pipeline diagram if you just want to see how a number gets made.
A sequel, not a first draft
My freshman project was an AI stock-recommendation engine — predictive models, risk profiles, testing against past market data. CareerStar is the same machine pointed at a harder, more personal question: rates stocks → rates careers. The idea that makes it work: a career isn’t simply “risky” or “safe.” What matters is whether the payoff justifies the risk — a field can be exposed to AI and still be worth pursuing if its growth, pay, and fit are strong enough, just as a risky stock can still belong in a portfolio.
I caught my own model overclaiming — and fixed it
So I rebuilt it for real. Every occupation now carries its genuine government skill profile — the U.S. Department of Labor rates each job on the same 68 abilities (35 skills + 33 knowledge areas) — and Fit measures how closely your interests line up with that profile, counting rare abilities more than common ones, because the skills every job shares (reading, critical thinking) would otherwise drown out the ones that actually make jobs different. When the app suggests “you might prefer this adjacent career instead,” that suggestion now really is the closest career by skill profile — not a guess.
The point isn’t that I wrote a bug. It’s that I went looking for where my own model was weakest, found it, and made the claim true instead of quietly deleting it. Don’t take my word for it — the full reconciliation document is committed in the repo, verdict by verdict.
Every number is math, not an LLM
One fixed formula is the single source of truth — the same inputs always give the same score, and no AI is involved in computing it. Each career’s growth and pay are ranked against all 730 occupations; risk combines published AI-exposure research with a stand-in for career instability; reward gets marked down by risk, and the result blends with personal fit:
The actual equations
Return = wGrowth·growth + wPay·pay Risk = wExposure·exposure + wVolatility·volatility RAV = Return · (1 − γ·Risk) Score = 100 · [ α·RAV + (1 − α)·Fit ]
The language model never computes a number. It only writes the plain-English “why,” from the already-computed score — and it never even sees your raw text, only controlled tags. The math is auditable; the AI is a narrator.
Back-tested against a real decade
The strongest evidence isn’t a formula — it’s history. I recovered the government’s original 2014 career forecasts from the Internet Archive, had the model score the 2014 labor market using only what was knowable then, and compared against what actually happened by 2024 — 647 occupations across a real decade. Careers it scored higher really did fare better, and it flagged 48% of the careers that went on to shrink where guessing would catch 33%. And the part a marketer would cut: the AI-risk penalty added nothing for that decade — exactly as it should, because 2014–24 ended before ChatGPT, and the AI-exposure number is a bet about the future that a test on the past can’t score. The misses are published by name, including the oil-crash casualties no career-by-career model catches.
Robust to ±20%, and the risk axis isn't redundant
Does the answer survive disagreement? The model’s weights are choices, so every comparison is re-run 729 times with each weight nudged 20% up and down in every combination. The app reports how often each career keeps its place, flagging a “close call” instead of selling a shaky #1 as certain.
Is the AI-risk number redundant? A fair objection is that “exposed to AI” might just be another way of saying “already declining.” It isn’t: across all 730occupations, knowing a career’s AI exposure tells you almost nothing about whether it’s projected to grow or shrink — the two are nearly unrelated in the data (correlation 0.075, where 0 means no relationship). So the risk number carries real information the growth forecast doesn’t, which is why the score adjusts for it instead of ranking on growth alone.
Five models, not one — because a formula is an opinion
The deepest critique of any scoring app is: who says your formula is right? My answer is to stop pretending there’s one formula. Every comparison is also scored under four rival models — an optimist that ignores AI risk entirely, a safety-first model that cares most about shelter from AI, one that asks how much reward you get per unit of risk, and a plain average of everything — and the app shows the spread. When all five agree, the ranking isn’t an artifact of my assumptions. When they crown different winners, the app says so, because that disagreement is the answer: your choice turns on how much AI risk you believe in, and no formula can settle that for you.
(The plain average is there on purpose: if a clever model can’t beat a simple average, its complexity isn’t earning its keep — a lesson borrowed from investing research.)
Solo, one summer, AI-native, on AWS
- One person, one summer — from brief to PRD to architecture to build, using a structured AI-native (BMAD) workflow.
- Real public data: U.S. BLS Employment Projections, O*NET 29.0 (skills + knowledge), Eloundou et al. AI-exposure, and College Scorecard for the education-ROI layer.
- Deterministic & stateless: no database, no accounts, no stored input — the whole model is a committed data file + pure functions.
- Deployed on AWS Amplify with CI/CD, HTTPS, security headers, rate limiting, and prompt-injection-resistant LLM use.