AI Career Setup Engine
CareerArc AI — bridging the gap between talent and ATS algorithms through automated, context-aware resume alignment.
Use the arrows inside the viewer to move through the slides, or open it full screen for a larger view.
Roughly 75% of resumes are filtered before a human reads them — screened out on keyword and formatting mismatches, not on capability. The steepest drop in the application funnel is the very first gate, and it is fully automated. Existing resume builders sell templates that optimise how a document looks; none compare it, line by line, against the job description in hand. The feedback candidates do get has no coordinates: 'add more keywords' names no terms, no weighting, and no placement.
- A live, public web product on Vercel, built on a Gemini-backed agentic pipeline for semantic matching and weighted keyword extraction.
- Generative STAR rewrites that turn a flat responsibility line into a quantified, JD-aligned achievement, with every recommendation traceable to the requirement that triggered it.
- A North Star metric — score-improving analyses per active user per week — chosen deliberately over raw scan volume, which rises when a product is confusing.
- A three-phase roadmap sequenced for retention, then distribution, then enterprise: a live editor, a Chrome extension and LinkedIn optimiser, and a B2B portal for coaches and placement cells.
- 01
Designed a two-input flow with zero configuration — upload a resume, paste the job description, run one analysis. No account, no profile to build, no stored uploads.
- 02
Built a five-stage pipeline: ingest unstructured PDF/DOCX and free-form JD text, parse and normalise both sides, score, map gaps by severity, then generate action.
- 03
Decomposed the score into four independent signals — ATS overall, keyword match, experience relevance, and structure/formatting — because a 40% keyword score and a 40% formatting score demand completely different work.
- 04
Resolved every weighted JD term into one of four states — exact match, synonym, partial overlap, or missing — so semantically equivalent phrasing still earns credit rather than failing a string comparison.
- 05
Sequenced the result screen as confidence → diagnosis → prescription, ranking gaps by severity so a thirty-item list becomes three things to fix tonight.
- 06
Normalised both inputs before scoring to solve the hardest problem: two-sided variance, where the same candidate would otherwise get different scores against two postings for the same role.