Cadences: The AI-Native Project Management Platform We Built from Scratch

Most project management tools add AI as an afterthought — a chatbot in a sidebar that summarizes tasks. Cadences is different. We built the entire platform with AI at the core: multi-provider intelligence that creates projects from natural language, voice agents that make real phone calls, WhatsApp automation, IoT device control, visual workflow programming, and offline-first sync with conflict resolution. Not a wrapper around someone else's API — a full platform at cadences.app, running on the edge with 0ms cold starts.

This article is the overview of the ecosystem that all our other projects live inside. Every tool we've written about — Perspectiva Studio, Transcriptor, Heartbeat Engine — runs on Cadences infrastructure.

The AI Brain: 7+ Providers, One Interface

Cadences routes to the best AI provider for each task:

Provider	Strengths	Used For
Gemini 2.5	Large context, structured output	Project creation, analysis, reports
GPT-4o	Balanced quality	General assistance, content
Claude Sonnet 4	Reasoning, code	Technical analysis, complex logic
DeepSeek	Cost-efficient reasoning	Bulk processing, data analysis
Groq (Llama 4)	Speed	Low-latency tasks, quick responses
Workers AI	Edge, zero latency	Embeddings, classification
xAI	Reasoning	Complex queries, alternative perspective

Every AI call goes through our AI Gateway with automatic fallback, cost tracking, and rate limiting. If Gemini is down, it falls back to GPT-4o. If GPT-4o is expensive for a bulk task, it routes to DeepSeek. The user never sees the routing — they just get answers.

The AI isn't a chatbot. It's an agent with access to your data:

• "Create a project for the Q1 marketing campaign with 12 tasks" → actually creates it, with realistic deadlines
• "What's our task completion rate this month across all projects?" → queries real data, returns analysis
• "Draft a follow-up email for the 15 leads that haven't responded" → generates personalized emails with merge fields

8 Project Views: Every Way to See Your Work

View	What It Does
Kanban	Drag-and-drop columns by status, advanced filters
Calendar	Day/week/month with Google Calendar integration
List	Compact view with grouping and sorting
Timeline / Gantt	Time bars, dependencies, milestones
Data Table	Airtable-style — dynamic schema databases
Workflow Canvas	Visual automation editor with connectable nodes
Monitors	Real-time metrics dashboards
Daily Agenda	Today's summary view

Data Tables: Airtable Inside Your PM Tool

This is the feature Notion users wish they had with real project management. Any project can become a dynamic database where each task is a row with configurable schema:

• Field types: String, Number, Boolean, Date, DateTime, Email, URL, JSON
• Auto-generated forms: Dynamic CRUD forms from the schema
• Public forms: Embeddable external forms, no login required
• Global search + column filters
• AI natural language queries: "Show me all leads from Barcelona with budget over €5,000"
• Automated scraping: Populate Data Tables from external websites

Use cases: CRMs, product inventories, service catalogs, sales records, contact databases. Each one is a project with DATA_TABLE type.

Voice Agents: Real Phone Calls with AI

Cadences makes actual phone calls with AI voices. Not pre-recorded messages — real-time conversational AI over telephony:

• 15+ realistic Spanish voices via ElevenLabs
• Inbound and outbound calls via Twilio
• Real-time LLM conversation — the AI responds naturally to whatever the person says
• Automatic transcription of every call
• Workflow integration — calls can trigger or be triggered by workflow automations

A real estate agency can set up an inbound agent that answers calls 24/7, qualifies leads, checks property availability against the Data Table, and schedules viewings on the Calendar. All without human intervention.

WhatsApp Automation: Official + Stealth

Dual-mode WhatsApp integration:

Mode	Advantage	Trade-off
Official API (Meta)	Enterprise reliability, approved templates	Per-message cost, template approval process
Local Agent (Electron)	Zero per-message cost, human-like behavior, anti-detection	Requires desktop client running

Both modes support: text, images, audio, location, chat reading, message queues with retries. Connect either to workflows for automated follow-ups, appointment reminders, or marketing sequences.

Visual Workflow Engine

The Workflow Canvas is a visual programming environment where you connect nodes to build automations:

Trigger (new lead form submission)
  │
  ├── AI Node: Qualify lead (score 1-10)
  │
  ├── Condition: Score > 7?
  │     ├── Yes → Voice Agent: Call lead within 5 minutes
  │     └── No → WhatsApp: Send nurture sequence
  │
  ├── Data Table: Update lead status
  │
  └── Notification: Alert sales team

The workflow execution engine is 7,073 lines of JavaScript. It handles long-running workflows via Cloudflare Durable Objects, retry logic, conditional branching, loops, and parallel execution.

IoT Hub: Control Physical Devices

Cadences controls physical hardware integrated into workflows:

Protocol	Devices
MQTT	Sensors, actuators, smart home devices
Serial	Arduino, ESP32 — direct communication
HTTP REST	Device APIs, webhooks
RTSP/ONVIF	IP cameras — live feeds, motion detection

A workflow can read a temperature sensor (MQTT), check if it exceeds a threshold (condition node), turn on the AC (HTTP REST), and send a WhatsApp alert to the facility manager. Physical and digital automation in one canvas.

Offline-First Sync: Works Without Internet

Cadences works offline. Full functionality, stored in IndexedDB, syncs when connection returns:

• IndexedDB — local storage, works without any network
• Google Drive — cloud sync with conflict resolution
• Polling every 5 seconds with change detection
• Tombstones for distributed deletion
• Vector clocks for field-level merge
• Atomic transactions for consistency

Field workers, remote teams, areas with spotty connectivity — Cadences keeps working. Changes merge intelligently when sync resumes, with field-level conflict resolution instead of last-write-wins.

Technical Architecture

Frontend (React 18 + Vite + Tailwind)
├── App.jsx              ~5,187 lines — application core
├── WorkflowView.jsx     ~5,086 lines — visual workflow editor
├── 50+ main components
├── IndexedDB            offline-first storage
└── Google Drive sync

Backend (Cloudflare Workers — Edge Computing)
├── 60+ serverless API endpoints
├── D1 (distributed SQLite)   107+ migrations
├── R2 Object Storage         Media, PDFs, DICOM
├── Durable Objects           Long-running workflows
├── KV                        Cache, sessions
└── AI Gateway                Multi-provider routing

Integrations
├── Google Calendar/Sheets/Drive
├── Twilio (telephony)
├── ElevenLabs (voices)
├── Meta WhatsApp Business API
├── MQTT brokers (IoT)
└── 7+ AI providers

Measured Results

Metric	Improvement
Project creation time	-96% (natural language → full project)
Ticket processing time	-92% (AI classification + routing)
Lead conversion	+40–65% (automated follow-up sequences)
Admin task automation	Up to 80% (workflows replace manual work)

Cadences 3.0: The Modular Rewrite

We're in the middle of a major architectural evolution:

Before	After
Monolithic files (7,000+ lines)	Modules <300 lines each
MLContext.jsx (3,479 lines)	17 modular files (~150 lines avg)
Hardcoded AI providers	15+ providers with unified base interface
Low test coverage	Target >70%
Large bundle	Target <2 MB gzipped

The v3 rewrite follows a local-first AI, privacy-first ML, modular architecture vision. 18 ML module files and 7 ML Trainer feature files are already migrated. The provider interface pattern means adding a new AI provider takes one file, not a codebase-wide refactor.

Key Takeaways

1. AI-native means AI at the foundation, not the sidebar. When every feature can call AI — project creation, data queries, voice calls, workflow decisions, content generation — the whole product experience changes. Users stop describing what they want to do and start describing what they want to achieve.

2. Multi-provider AI routing eliminates vendor lock-in and maximizes value. Different providers excel at different tasks. Routing automatically based on task type, cost, and availability means you always get the best result for the price. When a provider has an outage, users don't even notice.

3. Offline-first sync is harder than it sounds, but worth every line of code. Vector clocks, tombstones, atomic transactions, conflict resolution at the field level — it's complex. But the moment a field worker uses the app in a dead zone and everything syncs perfectly when they're back online, you know it was worth it.

4. Voice AI + WhatsApp + Workflows create automation loops that run businesses. A lead fills a form, the workflow qualifies them with AI, calls them with a voice agent, follows up on WhatsApp, updates the CRM, and notifies the sales team. The entire lead-to-meeting pipeline runs without human intervention.

5. Edge computing on Cloudflare Workers is production-ready for complex apps. 60+ endpoints, 107+ D1 migrations, Durable Objects for long-running workflows, R2 for storage, KV for caching. Zero cold starts, global distribution, and the cost is a fraction of traditional cloud. The platform proves that Workers can handle real application complexity, not just edge functions.