AI SaaS · 2026 Live

Cue

Production-grade WhatsApp Business AI agent with structured tool use and a full audit trail.

Tech stack

Next.js TypeScript PostgreSQL Meta Cloud API Anthropic Claude Groq Vercel

The problem

Most WhatsApp bot tools stop at hard-coded keyword flows or shallow GPT wrappers with no audit. Regulated workflows (support, billing, customer onboarding) cannot ship without an answer to what the bot said to a customer at 3am, with token counts, latency, tool calls, and errors logged per run. Existing tools do not provide that level of observability.

Goals

Receive and send on the Meta Cloud API webhook end to end
Ground every reply against a real knowledge base with a bound search tool
Treat every AI reply as a first-class run with full telemetry persisted
Make escalation rules explicit, not buried inside prompt soup
Keep the model provider pluggable so Groq, Claude, or BYO key all work

The solution

Webhook endpoint on Vercel receives WhatsApp messages and persists them with conversation context
Structured tool loop with four bound tools (lookup customer, search KB, draft reply, escalate to human)
Lexical KB lookup today with the pgvector schema in place for the Phase 3 embedding pipeline
Outgoing messages flagged ai_drafted and ai_sent so human-vs-AI authorship stays traceable
Escalation tool flips conversation status and writes an escalation row, surfaced in the admin dashboard
Admin dashboard surfaces conversations, escalations, and AI run details including tokens and latency

My role

→ Solo architect and engineer, requirements to deploy
→ Meta Cloud API webhook integration on Vercel
→ Agent loop design with structured tool use against Groq and Claude
→ KB schema and the four bound tools wired into the agent loop
→ Admin dashboard for conversations, escalations, and run inspection

UI direction

Operator-first admin built for monitoring conversations and AI runs, not for end users. End users only see WhatsApp on their phone, which is the point.

User flows

Inbound message flow

1 Customer sends a WhatsApp message to the business number
2 Meta Cloud API webhook hits the Cue endpoint, message persists with conversation context
3 Agent loop runs: lookup customer, search KB, draft reply, decide on escalation
4 Reply sends back over the Meta Cloud API, flagged ai_sent
5 Every model run, tool call, and token count writes to the Postgres audit trail

Escalation flow

1 Agent calls escalate_to_human based on explicit rules
2 Conversation status flips to escalated and an escalation row writes in the same transaction
3 Operator picks up the conversation from the admin dashboard
4 Subsequent operator replies persist alongside the AI run history

Screenshots

Click any image to open at full size.

Conversation list in the admin dashboard with status, last message preview, and escalation flags.

Conversation timeline showing inbound and outbound messages alongside AI run metadata and tool calls.

Key learnings

Treating every AI reply as a logged run with tokens, latency, and tool calls is the difference between a demo bot and a regulated-workflow bot
Explicit escalation rules in code beat trying to coax escalation behavior out of a prompt
A pluggable provider interface (Groq default, Claude via BYO key) lets the demo stay free while paying clients still get Claude
Persisting AI run metadata at the message level (not at the conversation level) keeps the audit trail useful when an agent loop spans multiple tool calls
Phase 3 next: swap the lexical KB lookup for pgvector embeddings; Phase 4 next: Slack notification on escalate plus an escalation queue UI

Want something like Cue?

I'm open to senior contract work. Let's talk about what you're building.

Get in touch