AI-Powered Resume Builder with Real-Time Streaming

Impact Statement

A sophisticated AI resume editor featuring real-time streaming updates and surgical JSON patching for seamless collaborative editing.

Overview

A sophisticated AI-powered resume editor that enables real-time, chat-based resume creation and optimization. The system uses surgical JSON patching (RFC 6902) to apply incremental updates as the AI generates content, creating a collaborative editing experience similar to Google Docs.

Key Features

• Real-time streaming - Watch content appear as the AI generates it
• Surgical updates - Only modified fields update, preserving context
• Intelligent suggestions - AI understands resume best practices
• Artifact system - Side-by-side chat and resume preview

Technical Stack

• Frontend: Next.js 15 with React 19 and App Router
• AI Integration: Vercel AI SDK with multi-provider support
• Styling: Tailwind CSS 4 with Framer Motion animations
• Database: Postgres (Neon Serverless) via Drizzle ORM
• Authentication: Auth.js (NextAuth.js) v5 Beta
• Observability: OpenTelemetry with Vercel/SigNoz integration

Core Architecture

Surgical JSON Patching

Instead of replacing entire document content on each update, the system generates RFC 6902 JSON Patches that target specific fields. This enables:

• Minimal UI re-renders for better performance
• Precise scroll-to-edit behavior highlighting changes
• Efficient undo/redo operations

Deterministic Tool Governance

In multi-tool AI systems, overlapping capabilities create ambiguity. This project implements hard blocking at the code level:

• createDocument owns initialization
• patchResume owns resume modifications (surgical)
• updateDocument owns generic documents (text, code)

Streaming Infrastructure

The system leverages Vercel AI SDK’s partialObjectStream to emit partial JSON objects as the model generates tokens. Custom filtering ensures patches are only applied once values have stabilized.

Development Patterns

E2E Testing Strategy

Playwright tests are the source of truth, implementing:

• Synchronized readiness via body.hydrated markers
• Streaming completion detection through status attributes
• Animation vs. functional separation for reliable tests

Prompt Engineering

The AI uses structured prompts with:

• Dynamic context injection for resume sections
• Verbatim formatting rules for consistency
• Task aggregation to batch related updates

Lessons Learned

• Partial object streaming enables typewriter-style UX without custom infrastructure
• Tool governance at code level is more reliable than prompt engineering alone
• JSON Patch provides surgical precision that full-document updates cannot match
• End-state verification in tests is more stable than intermediate checks