BookPilot: Human Team vs. AI Agents in the Ultimate Test

When building modern multi-channel software, the complexity rises exponentially. A system that syncs inventory, parses images using machine learning, and manages active cloud infrastructure is typically the domain of dedicated, specialized engineering teams.

For our internal project BookPilot—a SaaS application designed to list and synchronize books across multiple global marketplaces simultaneously—we decided to run a benchmark: How does traditional human development time compare to our hybrid AI engineering workflow?

Here is the breakdown of what it takes to build BookPilot traditionally, and how our agentic AI workflow changed the math entirely.

1. Codebase Breakdown & Architecture

BookPilot is organized as a production-ready pnpm monorepo. Excluding assets and configuration files, the system consists of approximately 19,600 lines of typescript and TSX code:

• Mobile Client (apps/mobile): ~6,553 LOC
• Tech: Expo, React Native, Tailwind CSS (NativeWind).
• Features: Camera scanning with barcode OCR, local zustand state machines (useStore.ts), and seller integrations.
• REST API (apps/api): ~5,380 LOC
• Tech: NestJS backend, Prisma ORM, PostgreSQL.
• Features: Listing logic, webhook handshakes, and user billing integrations.
• Marketplace Connectors (packages/connectors): ~4,936 LOC
• Tech: SDK bindings, OAuth flows, and event listeners.
• Integrations: eBay, Shopify, WooCommerce, Amazon, AbeBooks, and Etsy.
• AI & Cloud Layer (packages/ai & packages/aws): ~1,295 LOC
• Tech: AWS Textract / OCR integration, Gemini Vision APIs.
• Features: Automated visual assessment of book condition and weight attributes from photos.
• Infrastructure (packages/infra): ~419 LOC
• Tech: AWS Cloud Development Kit (CDK).
• Features: Multi-AZ VPC, ECS Fargate services, RDS Database, and SQS FIFO queues.
• Shared Layer & Workers (packages/shared & apps/workers): ~966 LOC
• Tech: Shared Zod contract schemas, BullMQ background queues.

2. Staffing Estimations: The Traditional Route

Under standard software engineering conditions, building BookPilot from scratch requires multi-disciplinary expertise. We estimated three human-only resource scenarios:

For larger companies, Scenario C (Specialized Agile Team of 4-5 Developers) could shorten the timeline to 4 to 6 weeks. However, this increases overhead, requiring constant coordination, API contracts alignment, merge conflicts management, and deployment synchronization.

3. The AI Workflow: Building BookPilot in 10 Days

Instead of deploying a traditional team, we built BookPilot using a single full-stack developer equipped with our hybrid AI development workflow, which we explained in detail in our article on The Ultimate AI Development Workflow.

By combining ChatGPT Pro as the Conceptual Architect with Antigravity (powered by Gemini) as the Sandboxed Code Executor, we compressed months of development into just 10 days.

How the Hybrid Workflow Solved Key Engineering Challenges:

A. Multi-Channel Synchronization The Problem: Preventing double-selling of a single book. If a book sells on eBay, it must be locked instantly on Shopify and WooCommerce. AI Solution: The developer designed the queue logic in ChatGPT Pro. The resulting Zod schemas and BullMQ event handlers were then given to Antigravity. Antigravity mapped the packages, implemented the state machine, and verified type safety in the workspace.

B. AWS Infrastructure as Code (CDK) The Problem: Writing error-prone CloudFormation templates or CDK configurations. AI Solution: The developer requested a VPC, ECS Fargate task, and SQS FIFO setup. Antigravity generated the CDK typescript file, checked it against local AWS configurations, ran the compile test, and outputted a clean deployment script.

new FargateService(this, 'Api', {
  cluster,
  taskDefinition,
  desiredCount: 2
});

C. AI OCR & Gemini Vision Pipeline The Problem: Combining Textract raw OCR output with Gemini's visual analysis to classify book details (condition, ISBN, weight) with 95%+ accuracy. AI Solution: ChatGPT Pro structured the prompts. Antigravity integrated the @google/generative-ai package, wrote the client call wrappers, and executed local Jest tests to refine prompt inputs until the classification hit the targeted thresholds.

4. Key Takeaways

The BookPilot project proves that software engineering is undergoing a paradigm shift.

• Context is King: Because Gemini’s massive context window allowed Antigravity to read the entire BookPilot monorepo structure, there was zero drift between the mobile TSX client and the NestJS API models.
• Leveraging Sandboxes: The developer did not spend days debugging setup errors or compilation mismatches. The agent ran pnpm build in its sandbox, captured compilation logs, and resolved import issues autonomously.

By leveraging AI agents to manage execution and compilation, a single developer achieved the output of an agile team in a fraction of the time.

Plan your next project with us At Aopas, we specialize in high-performance web applications, mobile apps, and deep API integrations. We leverage the best AI tools to deliver robust, custom codebases at speed.

Speak with our software architects