How Step works

Most career advice is either generic blog posts or one-off mentor calls. Step is something else: a recommendation engine grounded in 10,500+ curated career paths across 76 countries, with honest evidence counts instead of made-up probabilities.

The dataset

We've curated 10,500+ career paths — real anonymized stories of how people moved from one role to another. Each path has the starting state (stage, field, role), the transition type (vertical promo, industry pivot, geo move...), the next role with timeframe, the 4–5 concrete actions that drove the transition, and the 24-month outcome.

The dataset spans 76 countries — US, UK, Italy, Germany, France, Spain, the Nordics, India, Canada, Australia, Singapore, Hong Kong, Japan, Korea, Brazil, Mexico, the broader LatAm, the UAE, Saudi Arabia and the wider MENA, plus sub-Saharan Africa and CIS. It covers elite tracks (FAANG SWE, MBB consulting, IB analyst, VC associate, Magic Circle law, Bain Capital PE) and underrepresented ones (NHS doctor pivots, Italian boutique consulting, bootcamp grads, dropout founders, returnship parents, military veterans, late-30s career switchers, neurodivergent ICs, FGLI students). Every path has a confidence label — high for well-trodden, low for selective or unusual.

The retrieval

When you submit your profile (stage, field, skills, education, past positions, languages, salary, 5-year vision, specific dilemma), we embed it with Voyage AI into a 1024-dimensional vector. We then retrieve the 5 most similar career paths from the dataset using cosine distance on pgvector. This isn't keyword matching — semantically similar paths surface even when the exact words differ.

The recommendations

We feed the retrieved paths plus your profile to Anthropic Claude (Haiku 4.5) with a tight prompt that enforces 14 hard rules: ground every recommendation in 1–3 retrieved paths, honor your priority order (position vs money vs location), respect education and past positions to differentiate, no invented statistics, salary realism anchored to your current comp, honest confidence labels.

Output is 3–4 ranked recommendations. Each carries:

• A leverage tag — foundation (without this, the vision is unrealistic), accelerator (compresses the timeframe), optional (useful but not gating). Hard cap: at most one foundation per plan.
• Path evidence — a literal count from the retrieved paths, e.g. “3 of 5 retrieved profiles took this exact action; 2 reached an equivalent outcome within 24 months.” Never invented percentages.
• 3–4 concrete 90-day actions — verb-led, time-bounded, replicable. Not “network more.”
• 12-month outcome — one specific sentence describing where you'd be.

Why no probabilities

Career advice tools love to say “83% match” or “+38% chance of promotion.” With 8,500 paths and no control group, those numbers would be invented. We refuse. Instead each move carries the leverage tag plus an evidence count from the retrieved paths — that's the honest version of “how much does this matter?”

What you also get

• Honest take — a 4–6 sentence paragraph from a senior peer voice. Direct where the data supports it, humble where it doesn't.
• What we don't know about you — gaps in your input that would change the recommendations if filled. Prompt for refinement.
• 3 closed follow-up questions — generated from your gaps, click an option to refine the plan in 30 seconds.
• Decision tree (Premium) — anchored on the foundation move: NOW → DAY 90 → MONTH 6 → MONTH 18 → YEAR 5, with branches and 3 outcome scenarios.
• Per-recommendation scenario expansion (Premium) — for each rec, the 3-month, 12-month, 5-year states plus risks and tradeoffs.

Privacy

Your inputs and generated plans are stored on our infrastructure (Supabase EU, eu-west-1). We don't share or sell your data. CV uploads are processed in-memory only and never persisted as files. Email magic-links and 90-day check-ins are opt-in.