Channelflow

The brief.

The stack.

The six modules.

• Centralized Inbox

  Three-pane layout: conversation list on the left, message thread in the center, AI suggestions and contact details on the right. WhatsApp and Instagram DM and comment threads land in the same place. AI messages are visually distinct from human replies at a glance. Staff see the AI's suggested next response — with source attribution and confidence — before deciding whether to use it, edit it, or take over entirely.

• Booking Management

  Day timeline view showing all reservations — whether the AI collected them from a chat conversation or staff created them manually. Conflict detection with inline resolution. Each booking block links directly to its originating conversation, so the full context is one click away. Week and list views for broader planning.

• Analytics Dashboard

  Response time trends, AI deflection rate, per-channel message volume, and late-reply breakdowns. Four KPI tiles compare current period against prior period with delta indicators. Headline chart breaks down who replied — AI vs. human — and when late replies occurred. Designed for non-technical managers: plain-language labels, no SLA jargon.

• Outbound Messaging

  Campaign composer for WhatsApp and Instagram outreach. Four-step wizard: pick channels, pick targets, compose, review. Supports WhatsApp Business pre-approved message templates with personalization variables. Per-recipient delivery status. Deep-link from each campaign recipient back to their inbox conversation.

• AI Agent Configuration

  Master on/off toggle that pauses AI across all conversations immediately. Knowledge base with file upload and URL import, with ready/indexing/failed status per article. Booking rules (party size limits, opening hours per weekday, blackout dates). Confidence threshold slider, tone settings, supported languages. A sandbox playground lets staff test the AI's behavior before activating it globally.

• Channel Connections

  Connect and manage WhatsApp Business and Instagram Business accounts. Per-channel health monitoring — healthy, warning, disconnected, paused — with inline reason display when something goes wrong. Volume stats and last-sync timestamps per channel. Team management with role-based channel scope: scoped members only see conversations from their assigned channels.

Architecture decisions.

• Webhook isolation keeps messaging reliable

  Hono.js runs as a separate process from the Next.js dashboard. The webhook handler receives a message, writes it to the database, and acknowledges to the provider in under 500ms — then hands the job to BullMQ. The AI processing loop never touches the webhook path, so a slow AI response or a dashboard deploy never causes a missed message.

• BullMQ for zero message loss

  Each inbound message is enqueued to Redis before the webhook returns 200. If the server restarts mid-processing, the job survives and resumes. Built-in retry with exponential backoff handles transient AI model errors without manual intervention. A dead-letter queue catches jobs that exhaust retries so they can be replayed from the dashboard.

• Multi-tier AI pipeline — right model for each job

  Not every message needs the expensive reasoning model. A fast, lightweight classifier runs first and returns a category and confidence score in roughly 200–300ms. Only messages that pass the guardrail check and meet the confidence threshold reach the full reply writer. This keeps cost proportional to complexity and keeps latency low for simple FAQ responses.

• Mastra for orchestration and observability

  Mastra acts as the AI conductor — it handles memory per contact, registers agent tools, manages the generate loop, and provides built-in input processors for prompt injection detection. Critically, its observability output feeds directly into the dashboard's AI Agent section, so the audit trail is a product feature rather than a developer debug log.

• Guardrails embedded in instruction, not a keyword table

  High-risk category rules — influencer collaboration requests, complaints, partnership pitches — are written in natural language directly in the agent instruction and enforced semantically by Mastra's input processors. A keyword table would let a rephrased request slip through; semantic enforcement catches the intent. Rules are updated in code, not from a dashboard form, which is intentional: these are core business constraints, not settings.

• Self-hosted removes per-message billing risk

  The client operates a VPS. Running on Coolify with Docker Compose means the entire stack — backend, frontend, database, queue, cache — lives on infrastructure the client controls. No per-message or per-seat SaaS billing on the platform layer itself. SSL via Let's Encrypt, auto-deploy from GitHub, health checks built into the Compose file.

What I'd do differently.

• Separation of concerns pays off at deploy time

  Isolating the webhook handler from the dashboard means each layer deploys and restarts independently. A dashboard redeploy never risks a webhook timeout window. This seems obvious in theory; shipping it as two separate containers made it a real constraint rather than a principle.

• Confidence scores beat hard routing rules

  The classifier's confidence output is more actionable than any keyword-based routing table. When confidence drops below threshold, the message routes to human — no edge-case maintenance. The circuit breaker makes the degradation path explicit: if the classifier goes down, everything routes to human until it recovers.

• The audit trail is a feature, not a log

  Logging every AI decision with its reasoning, model, category, and duration turned the agent_logs table into a product surface. Staff can look up why the AI said something specific, days after the fact. That visibility is what makes the system feel trustworthy rather than like a black box.

• Human handoff design matters more than AI accuracy

  The most important UX decision wasn't how good the AI replies are — it was how clearly the handoff is communicated and how fast a human can take over. The amber banner, the one-click takeover, and the 'Return to Agent' path collectively define whether staff trust the system enough to use it.