Capabl — AI Career Decision Simulator

Capabl turns your real evidence into an honest readiness score, then simulates career paths and their tradeoffs — so you reason your way to a decision. You choose; the AI never picks for you.

Comment

Inspiration

A college student staring at "what do I learn next to actually get hired?" faces overload: dozens of skills, conflicting advice, and no idea what moves the needle. Existing tools either oversimplify ("Learn React!") or overwhelm with 100-item skill checklists. Almost none help you reason through the tradeoffs of competing strategies. We built Capabl around one decision moment:

"Given my real evidence, which learning path do I commit to this month — and what am I giving up by choosing it?"

What it does

Capabl takes a student's real evidence — resume, projects, certifications, a mock interview, optional GitHub/LinkedIn — and a target role, then:

  1. Builds the role. An LLM reads the goal + resume and returns the 12 skills this exact role demands today, each with an importance weight summing to 100.
  2. Scores the evidence honestly. For each skill it measures five evidence sources and computes a readiness % using a fixed, auditable formula — never a number the AI made up.
  3. Simulates three real paths. For example, for one student profile: Quick Wins (+13%, ~2 weeks), Close Critical Gaps (+24%, ~3 weeks), and Balanced Sprint (+32%, ~4 weeks) — each generated by a different objective, so they're genuinely distinct, not one list reshuffled. Each shows the score gain, the effort, and the tradeoff (what you delay by choosing it).
  4. Hands the decision back to the student. Capabl never auto-picks. The student commits to one path, and it becomes a 7-day plan with free resources.

Confusion → clarity → action.

How we built it

  • Role intelligence & explanations: an LLM (Gemini 2.5 Flash) parses the open-ended goal and messy resume, infers required skills + weights, and writes the human-readable reasoning and weekly plan.
  • Evidence matching: semantic similarity using gemini-embedding-001 embeddings (cosine similarity), so "built a chatbot" correctly matches "LLM integration" despite different wording. Each skill score is the stronger of a structural keyword match and a semantic match.
  • The scoring engine (deterministic): every score, gain, and projection comes from a documented formula — not the LLM. Per skill:

$$readiness = 0.35 \cdot interview + 0.25 \cdot projects + 0.20 \cdot resume + 0.12 \cdot certifications + 0.08 \cdot roadmap$$

Role Match is the weight-blended sum of skill readiness; overall Career Readiness blends match, semantic fit, resume strength, profile, skill count, and ATS signals.

Our core design principle: LLM where meaning matters, rules where trust matters. The LLM never outputs a score, so it can't fabricate numbers — every figure a student sees is reproducible and explainable.

Why this needs AI: parsing an open-ended goal and free-text resume needs genuine language understanding (a rules engine can't infer which 12 skills a niche role demands); matching messy evidence despite different vocabulary needs embeddings (keyword matching misses real matches). But every number runs on deterministic rules — that separation is the difference between reasoning and hallucination.

Challenges we ran into

  • Keeping the three simulated paths genuinely distinct. We solved it by generating each from a different objective (closest-to-done, highest-weight unmet, best gain-per-effort) rather than re-sorting one list. We verified across 6 synthetic user profiles that paths come out distinct and gains rise coherently.
  • Drawing a hard line so the LLM never touches the math — it receives no scores and only writes explanations, which keeps every number reproducible.
  • Representing uncertainty honestly instead of handing students a false "correct answer" — solved with confidence bands that widen when evidence is thin.

Accomplishments that we're proud of

  • A decision tool that genuinely reasons through tradeoffs instead of generating a pros/cons list.
  • A fully auditable scoring pipeline where the AI never invents a number — rare for an LLM-powered tool.
  • A clean input → reasoning → output → decision → action pipeline that keeps the human in control at the one moment that matters: the choice.

What we learned

The hard part of a "decision tool" isn't generating options — it's representing uncertainty and tradeoffs honestly so a person can actually reason, instead of being handed a false answer. We also learned where AI genuinely earns its place (meaning) versus where it must stay out (trust-critical numbers).

What's next for Capabl — AI Career Decision Simulator

Wider role coverage, longer-horizon 30/60/90-day plans, and letting students re-run the simulation as their evidence grows to watch their paths shift over time.

Built With

Share this project:

Updates