Find out what your syllabus is missing before your students do.
Upload a bootcamp or university curriculum, pick a target role, and instantly see which skills employers are hiring for that your program doesn't cover — scored against real job postings by a deterministic, unit-tested engine.
Runs with zero API keys.
git clone && npm install && npm run devgives you the complete flow — bundled sample curricula, a corpus of ~50 realistic job postings, and the full deterministic analysis engine. Add an Anthropic key to upgrade skill extraction to AI, and a RapidAPI key to swap the corpus for live listings. Nothing is stubbed; demo mode runs the real engine on real text.
Start from a bundled sample (full-stack bootcamp, data-science bootcamp, or a university CS core) or drag-and-drop a PDF, DOCX, CSV, or TXT.
The suggested role is pre-filled from your sample. With a RapidAPI key it pulls live listings from LinkedIn, Indeed, Glassdoor and more; without one it serves a bundled corpus of realistic postings (clearly badged as sample data).
A coverage score, a prioritized learning roadmap (critical → recommended, with the companies asking for each skill), a demand-vs-coverage chart, categorized skill lists, and reproducible recommendations. Export as a Markdown report or JSON.
The analysis is hybrid — AI does the one thing it's uniquely good at, and everything measurable is deterministic and testable.
curriculum text ─┐
├─► skill extraction ──► canonical taxonomy ──► matching + scoring ──► gap report
job postings ───┘ (LLM or rules) (alias resolution) (pure functions)
-
Canonical skill taxonomy (
src/lib/skills/taxonomy.ts) — ~170 skills, each with alias resolution (js→ JavaScript,react.js→ React,k8s→ Kubernetes) and careful handling of the awkward cases: symbol-bearing names (C++,C#,.NET,Node.js), and the genuinely ambiguous single-letter languages (C,R,Go) which are only matched inside a technology list — a deliberate precision-over-recall choice. -
Deterministic extractor (
extract.ts) — a single boundary-aware pass with longest-surface-first span masking, soReact Nativewins overReactandNode.jsis never mistaken for a barenode. Same input, same output, every time. -
Matching & scoring engine (
engine.ts) — computes per-skill demand frequency, required-vs-preferred (from each posting's "nice to have" boundary), covered / missing / bonus classification, a demand-weighted coverage score, an evidence map, and the recommendations. No LLM in this path — every number traces back to a count you can see in the chart. -
Two extractors, one interface — with an Anthropic key, curricula are extracted by Claude via structured tool use with a strict JSON Schema (
llm-extract.ts) for better recall on prose-heavy syllabi, then reconciled back onto the canonical taxonomy. Without a key, the deterministic extractor handles both sides. Either way, the scoring is identical and deterministic. If the LLM call fails, it falls back to the deterministic path — the app never hard-depends on the API.
This split is the whole point: use AI for the fuzzy problem (reading intent out of messy text), and plain testable code for the exact one (counting, weighting, ranking).
| Layer | Technology |
|---|---|
| Framework | Next.js 16 (App Router) + React 19 |
| Styling | Tailwind CSS 4 + shadcn/ui |
| AI | Anthropic API — structured tool use for skill extraction |
| Job Data | JSearch via RapidAPI (optional) |
| Charts | Recharts |
| File Parsing | pdf-parse, mammoth, csv-parse |
| Tests | Vitest |
- Node.js 18+
That's it. API keys are optional (see below).
git clone <your-repo-url>
cd curriculum-gap-analyser
npm install
npm run devOpen http://localhost:3000 and click a sample — the full flow works with no configuration.
cp .env.example .env.local| Key | Unlocks | Fallback without it |
|---|---|---|
ANTHROPIC_API_KEY |
AI skill extraction from uploaded curricula | Deterministic taxonomy extractor |
RAPIDAPI_KEY |
Live job listings via JSearch | Bundled corpus of ~50 realistic postings |
The results dashboard always labels which mode produced it, so sample data is never passed off as live.
npm test34 unit/integration tests cover taxonomy integrity, extraction edge cases (boundaries, symbol-bearing and ambiguous languages), the scoring math, and an end-to-end run over the bundled corpus.
src/
├── app/
│ ├── page.tsx # 4-step wizard (input → search → analyze → results)
│ └── api/
│ ├── parse/route.ts # extract text from uploaded files
│ ├── jobs/route.ts # live JSearch, or bundled corpus fallback
│ ├── analyze/route.ts # run the hybrid analysis pipeline
│ └── samples/route.ts # sample curricula for the "try a sample" flow
├── components/
│ ├── curriculum-input.tsx # upload + paste + sample picker
│ ├── job-search-form.tsx # role/location search with source badge
│ ├── analyzing-view.tsx # staged progress (parse → extract → match → score)
│ ├── analysis-dashboard.tsx # scores, provenance, chart, export
│ ├── learning-roadmap.tsx # gaps sequenced by demand, with evidence
│ ├── skill-gap-chart.tsx # demand-vs-coverage bar chart
│ └── skill-list.tsx # color-coded skill badges by category
├── lib/
│ ├── skills/
│ │ ├── taxonomy.ts # canonical skills + alias index
│ │ ├── extract.ts # deterministic extractor
│ │ ├── engine.ts # matching, scoring, roadmap (pure)
│ │ ├── llm-extract.ts # Claude structured-tool-use extractor
│ │ └── *.test.ts # the test suite
│ ├── analysis.ts # orchestrator: choose extractor, stamp provenance
│ ├── report.ts # Markdown report generator
│ ├── jobs.ts # JSearch client + country detection
│ ├── parser.ts # PDF / DOCX / CSV / TXT extraction
│ └── demo/index.ts # bundled corpus loader + demo job search
└── data/
├── job-corpus.json # ~50 realistic postings across 6 roles
└── curricula.json # 3 sample curricula with realistic gaps
- Precision over recall for ambiguous tokens.
C,R, andGoare common English words. Rather than flag every "go to the store" as the Go language, the extractor only accepts them inside a delimited technology list. It's a deliberate trade — documented and tested — and the kind of edge case that separates a real matcher from a naive keyword scan. - The coverage score is defensible. It's the share of total employer demand (∑ frequency) that the curriculum covers — not an opaque AI judgment. You can reconstruct it by hand from the chart.
- Honest provenance. Every result is stamped with its mode (live/demo), extractor (AI/rule-based), and job source, and the UI surfaces all three.
MIT — see LICENSE.