paulprae.com

An AI career platform at paulprae.com. Recruiters chat with a Claude-powered career assistant, explore a structured resume, and request tailored resumes for specific roles — all grounded in real career data.

This project is both a product and a portfolio piece. The codebase demonstrates production AI engineering: streaming LLM integration, tool-calling, prompt engineering with grounding rules, security hardening, and a CI/CD pipeline that deploys with zero manual steps.

What It Does

Route	Purpose
/	AI chat — recruiter Q&A, resume downloads, tailored resume generation
/resume	Full resume with section navigation and PDF/DOCX/MD downloads
/api/chat	Streaming chat API with tool-calling

Architecture

Two independent workflows connected by committed data files:

Resume Pipeline (local)              Website + Chat (Vercel)
─────────────────────────            ──────────────────────────
LinkedIn CSV + knowledge JSONs       Next.js App Router
    ↓ npm run ingest                     ↓ npm run build
career-data.json ──────────────────→ /resume (static pre-render)
    ↓ npm run generate                   ↓
Paul-Prae-Resume.md                  /api/chat (streaming)
    ↓ npm run export                     │ system prompt = career data
PDF + DOCX                               │ + grounding rules
                                         ↓
                                     Claude Sonnet → SSE stream → UI

The pipeline and website are independent. Website development requires only Node.js. The chat API requires ANTHROPIC_API_KEY at runtime.

AI Chat System

The chat API (app/api/chat/route.ts) streams responses via Vercel AI SDK 6 with two tools:

• generate_tailored_resume — accepts a job description, calls Claude to produce a role-specific resume
• get_resume_links — returns download URLs for PDF, DOCX, Markdown, and web formats

Career data (~90K tokens) is loaded into the system prompt with Anthropic prompt caching (5-min TTL). After the first request, subsequent turns reuse the cached prompt at ~90% cost reduction.

System prompts include grounding rules (G1-G10) that constrain the model to only cite verified career data, and security rules (S1-S5) that defend against prompt injection. Prompts live in lib/prompts/ as Markdown files with YAML frontmatter.

Resume Pipeline

The pipeline generates career artifacts from structured data:

1. Ingest — parse LinkedIn CSV exports + knowledge base JSONs into career-data.json
2. Generate — call Claude Opus 4.6 with career data + brand guidelines → staging markdown resume
3. Compare — optional interactive section-by-section review (--judge for LLM scoring)
4. Approve — promote staging resume to approved (the version the website reads)
5. Export — Pandoc + Typst convert → PDF and DOCX

Generation writes to a staging file; the website reads the approved file. This prevents regeneration from overwriting reviewed content. All pipeline steps skip automatically when outputs are newer than inputs.

Tech Stack

Layer	Technology	Version
Framework	Next.js (App Router, TypeScript)	16.x
AI Chat	Vercel AI SDK (ai + @ai-sdk/anthropic) + Claude Sonnet 4.6	6.x
Chat UI	@assistant-ui/react (Radix-style primitives)	0.12.x
Styling	Tailwind CSS	4.x
AI Generation	Anthropic Claude API (Opus 4.6) for resume pipeline	—
Rate Limiting	Upstash Redis (@upstash/ratelimit) with in-memory fallback	—
Validation	Zod	4.x
Resume Export	Pandoc (MD→DOCX) + Typst (MD→PDF)	—
Testing	Vitest + Testing Library + Playwright	—
Linting	ESLint 9 + Prettier + husky + lint-staged	—
Deployment	Vercel via GitHub Actions CI/CD	—

Getting Started

Website Development (no API key needed)

git clone https://github.com/praeducer/paulprae-com.git
cd paulprae-com
npm install
npm run dev     # localhost:3000 with Turbopack hot-reload

The website reads committed data files and works immediately. For UI/style changes, this is all you need.

Full Pipeline Setup

To regenerate resume content from LinkedIn data:

cp .env.local.example .env.local
# Add your ANTHROPIC_API_KEY (https://console.anthropic.com/settings/keys)

# Place LinkedIn CSVs in data/sources/linkedin/
# Export from: https://www.linkedin.com/mypreferences/d/download-my-data
# Select "Download larger data archive" for full position descriptions

# Install export dependencies (for PDF/DOCX)
# Ubuntu/WSL: sudo apt-get install -y pandoc && cargo install typst-cli
# macOS: brew install pandoc typst

npm run pipeline     # ingest → generate → export

Commands

# Development
npm run dev                # Dev server (Turbopack)
npm run build              # Production build
npm test                   # 400+ unit/component tests (~500ms)
npm run test:e2e           # Playwright E2E smoke tests
npm run check              # Full pre-push checklist (lint + format + test + build + validate)

# Pipeline
npm run pipeline           # Full: ingest → generate → export
npm run pipeline:content   # AI steps only: ingest → generate
npm run ingest             # Parse LinkedIn CSVs + knowledge JSONs → career-data.json
npm run generate           # Claude API → Paul-Prae-Resume.staging.md
npm run export             # Pandoc + Typst → PDF + DOCX

Playwright notes:

• Default npm run test:e2e runs fast mocked smoke tests on Chromium only.
• Use PW_FULL_MATRIX=1 npm run test:e2e for Firefox/WebKit/mobile matrix.
• Use E2E_LIVE_CHAT=1 npx playwright test e2e/live-chat.spec.ts for optional real API validation.

Security

The chat API includes multiple defense layers documented in SECURITY.md:

• Origin validation — proxy.ts blocks cross-origin requests from unauthorized domains
• Rate limiting — 20 req/min per IP via Upstash Redis (in-memory fallback when Redis unavailable)
• Input validation — request body size (100KB), message count (50), per-message length (4K chars)
• Prompt injection defense — security rules S1-S5 in all system prompts, XML delimiters around user input in tool calls
• Security headers — CSP, HSTS, X-Frame-Options (DENY), Permissions-Policy via vercel.json

Deployment

Deploys through GitHub Actions — Vercel Git integration is not used:

Push to main → CI: lint, format, test, build, validate
             → Deploy: preview → smoke test → promote to production

Environment	URL	Trigger
Local	localhost:3000	npm run dev
Preview	*.vercel.app	CI passes → deploy workflow
Production	paulprae.com	Smoke test passes → auto-promote

AI generation runs locally. Vercel only runs next build against committed files — no API key needed in the build step.

Project Structure

app/                        Next.js App Router pages and layouts
  api/chat/route.ts         Streaming chat API with tool-calling
  components/               ChatHome (shared by / and /tools), QuickActions
  resume/                   Resume page with section nav, downloads
  tools/                    Job search content tools (noindex)
lib/
  agent/context.ts          Career context builder for system prompts
  prompts/                  System prompts (Markdown + YAML frontmatter + few-shot examples)
  data-utils.ts             Shared utilities (stripEmpty for token optimization)
  constants.ts              Shared constants (client + server)
data/
  sources/linkedin/         LinkedIn CSV exports (gitignored)
  sources/knowledge/        Knowledge base JSONs (committed)
  generated/                Pipeline outputs (career-data.json, resume .md committed; PDF/DOCX gitignored)
scripts/                    Pipeline scripts + Typst stylesheet
tests/                      Unit, integration, component tests (Vitest + Testing Library)
e2e/                        Playwright E2E smoke tests
proxy.ts                    Next.js 16 proxy (CORS + origin validation)
docs/                       Technical documentation

Documentation

Document	Purpose
docs/technical-design-document.md	System architecture, constraints, and phase roadmap
docs/ai-architecture.md	AI architecture decisions and well-architected review
docs/devops.md	Deployment, smoke tests, rollback, CI/CD
docs/uat-checklist.md	Manual QA checklist for post-deploy verification
docs/domain-dns-runbook.md	DNS operations, validation, troubleshooting
SECURITY.md	Security policy, threat model, cost controls
CHANGELOG.md	Release history
CONTRIBUTING.md	Development workflow and code standards

Roadmap

Phase 3: Knowledge-Graph-Augmented AI (planned)

• Neo4j knowledge graph (Person → Role → Company → Project → Skill → Outcome)
• AI agents with tool-calling via Claude Agent SDK
• n8n automation workflows for data ingestion and enrichment

License

MIT