baubar

Project management for construction — built for the office and the field.

The web app gives office teams a structured interface to manage projects, reports, companies, and contacts. Field workers will soon interact with the same data through WhatsApp, using an AI assistant that understands natural language and can look things up or create entries on their behalf.

Stack

• Next.js 14 (App Router) — web UI and REST API
• Supabase — PostgreSQL, auth, file storage
• Drizzle ORM — type-safe DB access
• Turborepo — monorepo build system
• Vercel AI SDK + Google Gemini — AI assistant and tool calling
• Hono — lightweight HTTP server for the agent service

Monorepo structure

apps/
  web/        # Next.js app — UI + REST API (/api/v1/...)
  agent/      # AI agent service (port 3001) — LLM, tools, conversation threads
  gateway/    # Messaging gateway — WhatsApp and future channels (stub)
packages/
  ai/         # Shared AI tools, system prompt, agent runner
  db/         # Drizzle schema + Supabase client
  core/       # Zod schemas, mutations, event emitter
  ui/         # Shared shadcn/ui components
  config/     # Shared tsconfig / eslint / prettier

How it fits together

Browser  →  apps/web  ──────────────────────────────→  REST API
                │                                           ↑
                └→  apps/agent (LLM + tools)  ────────────┘
                         ↑
WhatsApp  →  apps/gateway  ┘

The agent is channel-agnostic — the same LLM and tools power the web chat bubble today and will power the WhatsApp integration tomorrow.

Getting started

Prerequisites

• Node.js 20+
• Supabase project
• Google AI API key

Setup

npm install
cp apps/web/.env.example apps/web/.env.local
cp apps/agent/.env.example apps/agent/.env.local

Run the SQL migrations from CLAUDE.md against your Supabase project, fill in the env files, then:

npm run dev --workspace=apps/web    # port 3000
npm run dev --workspace=apps/agent  # port 3001

Environment variables

apps/web/.env.local

NEXT_PUBLIC_SUPABASE_URL=
NEXT_PUBLIC_SUPABASE_ANON_KEY=
DATABASE_URL=
AGENT_SERVICE_URL=http://localhost:3001

apps/agent/.env.local

PORT=3001
SUPABASE_URL=
SUPABASE_SERVICE_ROLE_KEY=
DATABASE_URL=
GOOGLE_API_KEY=
GROQ_API_KEY=
WEB_API_BASE=http://localhost:3000

LLM Evals

The evals/ folder contains a Promptfoo test suite for the AI agent. Tests run against real LLMs with a mocked API backend (fixture data, no database needed).

Run evals

# Run all tests in the terminal
npm run eval

# Run and open the browser UI
npm run eval:ui

# Open the browser UI for the last run (without re-running)
npx promptfoo view

Browser UI

npm run eval:ui (or npx promptfoo view) opens http://localhost:15500 with:

• Side-by-side comparison table across all tested models
• Pass/fail breakdown per test and per assertion
• Full model outputs for each test case

Adding a test

Tests live in evals/promptfooconfig.yaml. Each test has a message (the user turn), optional history (prior conversation turns to simulate mid-conversation state), and assert checks.

- description: what this test verifies
  vars:
    message: "User message in German"
  assert:
    # Check a tool was called
    - type: javascript
      value: "JSON.parse(output).toolsCalled.includes('list_projects')"
    # Check text content (single-line — no return needed)
    - type: javascript
      value: "output.toLowerCase().includes('mehrhaus')"
    # Multi-line assertions need explicit return
    - type: javascript
      value: |
        const { toolsCalled } = JSON.parse(output)
        return toolsCalled.includes('list_projects') && toolsCalled.includes('create_report')
    # Case-insensitive substring match in the text reply
    - type: icontains
      value: "Tagesbericht"

The output variable in JavaScript assertions is a JSON string. Parse it to access:

• JSON.parse(output).text — the model's text reply
• JSON.parse(output).toolsCalled — ordered array of tool names called (e.g. ["list_projects", "create_report"])
• JSON.parse(output).toolArgs — last args per tool name

Testing multiple LLMs

Add entries to the providers list in promptfooconfig.yaml. Every test runs against every provider and results appear side by side in the UI.

providers:
  - id: "file://provider.ts"
    label: "Gemini 2.5 Flash"
    config:
      provider: google
      model: gemini-2.5-flash
  - id: "file://provider.ts"
    label: "Groq Llama 3.3 70B"
    config:
      provider: groq
      model: llama-3.3-70b-versatile

Supported providers: google (requires GOOGLE_API_KEY), groq (requires GROQ_API_KEY), anthropic (requires ANTHROPIC_API_KEY + npm install @ai-sdk/anthropic).

Fixture data

evals/fixtures.ts contains the hardcoded projects, reports, and stubs returned by the mock backend. Edit these to match realistic test data — the mock intercepts all http://mock/api/v1/... calls so no real database is involved.