MenuPulse AI

Production-oriented full-stack portfolio project for restaurant delivery intelligence. MenuPulse turns channel, competitor, and menu performance data into validated recommendations and draft experiments through a conversational GLM copilot.

The interface was designed from an original visual concept and implemented as a responsive Vue application. The source concept is preserved in design/dashboard-concept.png.

Why this project

This repository demonstrates the exact combination required for an AI Full Stack Engineer:

• Vue 3 + TypeScript product interface
• Node.js + TypeScript backend
• PostgreSQL relational data model with Prisma
• OpenAI-compatible integration using Z.ai GLM through OpenRouter
• Structured outputs validated with Zod
• Prompt-injection checks, strict request schemas, rate limiting, CORS restrictions, security headers, and fallback behavior
• Deterministic demo mode, API tests, Docker Compose, and GitHub Actions

Product workflow

1. A restaurant operator reviews channel and menu performance.
2. The copilot receives a natural-language business question.
3. The API sends compact restaurant evidence to GLM.
4. The answer is parsed into a strict recommendation schema.
5. Invalid output is repaired once, then falls back safely to a deterministic fixture.
6. The operator converts the recommendation into a draft experiment.

Architecture

Vue dashboard
   │
   ├── GET /api/dashboard
   ├── POST /api/copilot
   └── POST /api/experiments
          │
       Fastify API
          ├── TypeBox request validation
          ├── injection / abuse checks
          ├── GLM provider adapter
          ├── Zod structured-output validation
          └── Prisma → PostgreSQL

Run locally

Requirements: Node 20+, pnpm, Docker.

cp .env.example .env
docker compose up -d
pnpm install
pnpm db:generate
pnpm db:migrate
pnpm db:seed
pnpm dev

Open http://localhost:5173.

The app starts in DEMO_MODE=true, so it works without an API key. To use GLM:

DEMO_MODE=false
LLM_API_KEY=your_openrouter_key
LLM_MODEL=z-ai/glm-5.1

Keep secrets local. Never commit .env.

Reliability design

• Input body limits and strict JSON schemas
• OpenRouter/Z.ai model is configurable rather than hard-coded in business logic
• Grounded prompt receives only dashboard evidence
• Low-temperature generation
• Zod validation for every model field
• One repair attempt with validation feedback
• Deterministic fallback when provider or schema validation fails
• Suspicious instruction override patterns return HTTP 422
• Recommendation confidence and evidence are always visible to the user
• Consequential action creates a draft experiment; it does not silently launch one

Commands

pnpm build
pnpm test
pnpm typecheck
pnpm lint

Roadmap

• Auth and organization tenancy
• Real delivery-platform ingestion workers
• Experiment result tracking
• Redis response caching and distributed rate limits
• Evaluation dataset for recommendation faithfulness and business impact
• Image-generation workflow for menu photography

Disclaimer

This is an independent portfolio project using synthetic data and generic channel names. It is not affiliated with Pleez, Uber Eats, Glovo, or Z.ai.