Boundless

Boundless — Resume intelligence, honestly scored.

Upload a resume. Paste a job description. Get a real match score — parsed with pdf.js, scored by a deterministic offline NLP engine, audited against 16 ATS rules, and optionally sharpened by AI with your own OpenRouter key.

v2 — the redesign. The original main branch was a beautiful static prototype: the UI existed, the engine did not — every score was hardcoded. This branch is the rework: a real document pipeline, a real scoring engine, a real backend, and an optional AI layer. The GitHub Pages demo of the prototype remains untouched on main.

What it does

Extract — server-side parsing of PDF (via unpdf / pdf.js), DOCX (via mammoth), TXT/MD, or pasted text. Section headers, date ranges, degrees, contact points and quantified achievements are identified structurally.
Understand — the offline engine tokenizes and stems both documents, runs TF-IDF salience over the job description, and matches skills through a 450+ entry taxonomy with alias-, stem- and fuzzy-matching (Jaro–Winkler ≥ 0.92), plus conservative implication rules (PostgreSQL experience satisfies a “SQL” requirement; GitHub Actions satisfies CI/CD).
Score — four weighted dimensions compose the match score, and 16 deterministic ATS rules audit parseability, formatting and keyword coverage separately:

Dimension	Weight	Signals
Skills	35%	required ×3 / preferred ×1 weighted coverage, mention-weighted
Keywords	25%	TF-IDF salient-term coverage + stem-level cosine similarity
Experience	25%	detected years vs the role’s seniority band
Education	15%	degree band vs the JD’s stated expectation

The ATS audit checks contact discoverability, standard sections, machine readability, length, quantified achievements, action verbs, bullet ratio, date consistency, title alignment, keyword density, skills section, education detail, pronoun discipline, extraction artifacts and location — each with a pass/warn/fail verdict, a weighted score and a concrete fix.

(Optional) Enhance — bring your own OpenRouter key and any model on the platform (Claude, GPT, Gemini, Llama, DeepSeek…) to get gap narratives, an ATS interpretation, a tailored summary rewrite and stronger bullet suggestions. The key lives in your browser’s localStorage, is forwarded only with your own requests, and is never persisted server-side. The AI’s structured JSON output is validated with zod before display — and if it fails, your offline results remain fully intact.

Architecture

src/
├─ app/
│  ├─ page.tsx                  # screen machine: landing → input → processing → results
│  ├─ layout.tsx                # Fraunces + Geist fonts, themes, toaster
│  └─ api/
│     ├─ analyze/               # POST — parse doc → run engine → persist → result JSON
│     ├─ enhance/               # POST — BYOK OpenRouter second pass (zod-validated)
│     ├─ history/ (+[id]/)      # GET/DELETE — analysis history (SQLite/Prisma)
│     └─ openrouter/test/       # POST — validate the user's key against /api/v1/key
├─ lib/
│  ├─ analysis/                # THE ENGINE (deterministic, zero network)
│  │  ├─ taxonomy.ts            # 450+ skills, 17 categories, alias-aware
│  │  ├─ text.ts                # tokenizer, Porter stemmer, stopwords, n-grams
│  │  ├─ sections.ts            # resume section detection
│  │  ├─ resume-parser.ts       # contact, dates, degree, metrics, years of experience
│  │  ├─ jd-parser.ts           # title, seniority, required vs nice-to-have split
│  │  ├─ matcher.ts             # alias + stem + Jaro-Winkler + implication matching
│  │  ├─ tfidf.ts               # TF-IDF salience + cosine similarity
│  │  ├─ ats-rules.ts           # 16-rule weighted audit
│  │  ├─ scoring.ts             # dimension scores, overall, recommendations
│  │  └─ index.ts               # defensive orchestrator
│  ├─ parse-document.ts         # unpdf / mammoth / text extraction
│  ├─ openrouter.ts             # prompt builder, BYOK client, zod schema
│  ├─ store.ts                  # zustand settings (key, model, mode) + samples
│  └─ db.ts                     # Prisma client
└─ components/
   ├─ ui/                       # shadcn/ui (New York)
   └─ boundless/                # landing, input, processing, results, AI panel,
                                # settings dialog, history drawer, chrome, primitives

Stack: Next.js 16 (App Router) · TypeScript · Tailwind CSS 4 · shadcn/ui · Prisma + SQLite · Framer Motion · Zustand · Zod · sonner · natural (NLP) · unpdf/pdf.js · mammoth.

Getting started

# 1. install
npm install        # or bun install / pnpm install

# 2. environment
cp .env.example .env
# DATABASE_URL=file:./dev.db  (SQLite — no server needed)

# 3. database
npm run db:push

# 4. run
npm run dev       # → http://localhost:3000

Optional: verify the offline engine determinism — npm run engine:check runs the engine’s self-check (positive and negative controls, sanity assertions).

Notes

Privacy: resumes are parsed in server memory and stored in the local SQLite database only (for the in-app history). Nothing is sent anywhere unless you enable the AI layer with your own OpenRouter key.
Branching: main preserves the original static prototype (deployed on GitHub Pages). redesign is this full-stack rework.

Aligning your resume with endless possibilities.