Inspiration
USAII hackathon
What it does
Most decision tools fail in one of two ways — they oversimplify into a pros/cons list, or they overwhelm you with forty tabs of research. Neither one actually helps you think.
Pathwise sits in between. You pick two real paths — say, grad school versus a full-time job — and tune the numbers to your actual situation: program cost, starting salary, how much runway you've got. The engine scores both paths across four lenses — money, growth, stability, and personal fit — at three points in time: one year out, three years out, ten years out. And every single number comes with a confidence level, shown right there in the chart, because pretending to know what your career looks like in 2036 would just be dishonest.
Then it surfaces the stuff a list never would — like the fact that a startup's real cost isn't on a balance sheet, it's the years you can't get back. And it generates a plain-language read of all of it, either from a built-in template or, if you want, a live Claude call.
The one thing it refuses to do is tell you what to pick. That's not a missing feature — that's the point.
How we built it
Challenges we ran into
codes, API and working model
Accomplishments that we're proud of
deployed, working and available for everyone to use
What we learned
(1) Inputs: Two user-selected life/career paths (from grad school, full-time job, startup, bootcamp), user-edited numeric parameters specific to each path (program cost, funding %, starting salary, growth rate, runway, etc.), four priority weights (financial / growth / stability / fit, 1–10 each), optional free-text description of the user's situation, and optional context flags (dependents, location flexibility).
(2) AI capability used: A two-layer hybrid. Layer 1 is a deterministic decision-support / reasoning-chain system: rule-based tradeoff scoring across a 4-lens × 3-horizon matrix, with explicit confidence bands, plus rule-based detection of hidden considerations that fire based on path-combination and context. Layer 2 is an optional generative NLP layer (Claude) used exclusively to translate the already-computed structured scores into plain-language explanation — it is explicitly prompted not to alter or override those scores. A template-based fallback narrator covers this layer when no API key is supplied, so the app is fully functional without any LLM call.
(3) What processing happens: The rules engine computes per-lens/per-horizon scores from the user's specific parameters (not a static lookup), applies horizon-weighted averaging using the user's stated priorities to compute a "lean," runs a set of contextual rules to surface hidden considerations specific to the chosen path combination, and (optionally) sends the structured result to an LLM for narration only.
(4) Outputs — what the user receives: A two-path comparison chart (4 lenses × 3 time horizons, with visible confidence indicators), a weighted "lean" (explicitly not a recommendation — see Human-in-Loop below), a list of hidden considerations specific to their situation, optional what-if scenario deltas, and a plain-language narrative explanation.
What's next for Pathwise
being judged and used by other person to tell their experience with it
