Core full-stack engineer on ClientServe AI at HashThink — NestJS multi-tenant API, AI voice / telephony, campaign orchestration, Meta and CRM automations, and production dashboard features — working alongside the product team on a live SaaS at clientserve.ai.

Contributions

Multi-tenant workspace model with isolated Twilio subaccounts, numbers, and billing
Inbound / outbound AI voice path: Twilio Media Streams ↔ OpenAI Realtime ↔ ElevenLabs TTS
Meta Lead Ads OAuth, page picker, webhook leadgen → CRM → optional auto-outbound call
Instagram and Messenger social automation (connect, webhooks, lead capture, auto-reply)
Campaign orchestrator with sharded dial queues, audiences, retries, and run analytics
Missed-call retry and booking-reminder job pipelines (BullMQ)
Google Calendar booking tools for the live voice agent
Stripe Checkout, webhooks, and metered voice usage
Knowledge pipeline: website crawl, document embeddings, Knowledge Hub → agent tools
Production ops: EC2 / Nginx / PM2 deploys, Sentry, staging vs production API surfaces

1. Real-time Telephony + AI Audio Bridge

Challenge: Keeping low-latency, duplex audio between Twilio Media Streams, OpenAI Realtime, and ElevenLabs TTS without dropouts, echo, or stalled sessions under variable network conditions.

Solution: Built a dedicated voice session layer that bridges Twilio streams to Realtime tool-calling sessions, streams ElevenLabs µ-law audio into the call, and handles reconnect / teardown cleanly per call lifecycle.

2. Multi-Tenant Telephony Isolation

Challenge: Agencies need many clients on one product without shared numbers, call history, or billing bleed.

Solution: Modeled Workspace as the hard tenancy boundary — per-workspace Twilio subaccounts, numbers, Trust Hub rows, Meta configs, CRM data, and Stripe usage events — with active-workspace preference for the dashboard.

3. Speed-to-Lead Automation Under Webhook Noise

Challenge: Meta Lead Ads and WhatConverts fire noisy, duplicate, or incomplete payloads; auto-calling the wrong or empty phone wastes money and damages trust.

Solution: Normalized lead field parsing, idempotent customer upserts by lead IDs, phone validation before dial, and feature flags for auto-outbound so workspaces can enable callbacks only when ready.

4. Campaign Dialer Fairness and Throughput

Challenge: Concurrent outbound campaigns across tenants can starve each other or hit Twilio concurrency limits.

Solution: BullMQ coordinator with dial shards hashed by auth user, configurable shard count and concurrency caps, per-run attempt / retry / stop rules, and outcome logging on campaign runs and calls.

5. Knowledge That Stays On-Brand

Challenge: Generic LLM answers invent hours, pricing, and services that hurt conversion and compliance.

Solution: Website crawler + document chunk embeddings (pgvector-style) with Realtime search_knowledge_base tools and Knowledge Hub prompt sync so the agent prefers business-sourced facts.

6. Human-First Call Forwarding with AI Fallback

Challenge: Businesses want staff to answer when available, but still need AI when the office misses the call.

Solution: Twilio Dial to the business line first; on no-answer / busy outcomes, dial-action webhook connects the same caller into the AI Realtime session without forcing a second customer callback.

7. Meta Ads Surface vs Instagram-Only Promotion

Challenge: /connect-ads is Page-centric; Instagram-only promotions without a linked Facebook Page do not appear in Page-filtered campaign lists.

Solution: Documented Page-linked matching (promoted_object.page_id) as the supported path; Instagram social automation remains a separate connect flow for DMs/comments; future work can match instagram_user_id or offer ad-account-level campaign views.

8. Billing That Matches Voice Cost Reality

Challenge: Voice minutes vary wildly by geography and call length; flat SaaS pricing alone underprices heavy dialers.

Solution: Stripe subscriptions plus metered usage events for voice, plan gates on CRM / chatbot / social features, and clear Professional vs Agency packaging for single businesses vs multi-workspace operators.

9. Post-Call Continuity

Challenge: A completed call without a written trail forces staff to re-listen to recordings.

Solution: Persist transcripts and recordings (S3), optional delayed SMS summary to the customer, and CRM / WhatConverts follow-up sync phases after outbound automation.

10. Production Reliability on a Lean Ops Stack

Challenge: Single-EC2 production with memory-sensitive Node processes must survive deploys, spikes, and long-running Media Stream connections.

Solution: PM2 process separation for API and frontend preview, memory restart thresholds, Nginx reverse proxy, Sentry error tracking, and staging (test.clientserve.ai / test-api.clientserve.ai) for safer rollout.

1. Tenancy First

Workspace isolation must be designed into schema, Twilio provisioning, and OAuth tokens from day one
Agency products fail quietly when one client’s Meta page or number leaks into another workspace

2. Voice Is a Systems Problem

LLM quality is only one slice — telephony codecs, stream lifecycle, and tool latency dominate perceived quality
Always design for hangups, mid-call tool failures, and partial transcripts

3. Webhooks Need Idempotency

Ad platforms retry; lead forms send test payloads; phone fields are messy
Upsert by external lead IDs and gate auto-dial behind validated phones

4. Async Queues Protect the Request Path

Campaign dials, crawls, retries, and Stripe usage belong on BullMQ, not in HTTP request threads
Shard and concurrency limits are product features, not just infra knobs

5. Integrations Are Product Surfaces

“Connected” is not enough — page pickers, ad account filters, reconnect states, and empty-state copy determine whether agencies trust the Ads tab
Separate social OAuth (IG/Messenger) from ads OAuth when Meta scopes and review differ

6. Knowledge Beats Prompting Alone

RAG + crawl beats a long static system prompt for hours, menus, and service areas
Sync Knowledge Hub into the live agent so dashboard edits show up on the next call

7. Metered Billing Aligns Incentives

Outbound campaigns can burn minutes fast; usage metering keeps unit economics honest
Plan gates communicate which automations are Professional vs Enterprise / Agency

8. Ops Simplicity Scales Further Than Expected

EC2 + Nginx + PM2 is enough for early production if process boundaries and monitoring are clear
Staging API/UI parity prevents “works on Lovable, breaks on prod” surprises

Impact Highlights

Faster first contact on paid leads vs manual dial queues
After-hours and missed-call coverage without staffing a night shift
Fewer unqualified calendar bookings via voice qualification + Google Calendar tools
Agency ability to stand up a new client workspace with isolated numbers and ads
Target speed-to-lead: contact new ad leads in under 60 seconds of webhook receipt
Coverage: inbound + outbound AI voice, Meta Lead Ads, Instagram / Messenger, Google Calendar, HubSpot, Square, WhatConverts

ClientServe AI is a full-stack SaaS with a NestJS 11 API and a Vite + React 18 dashboard SPA. Tenancy is centered on Workspace: nearly all domain models carry workspaceId, and the UI switches active workspace for agencies managing multiple clients.

The backend is modular (Auth, Twilio, AssistantVoice, Campaigns / CampaignOrchestrator, Meta, HubSpot, Square, WhatConverts, Calendar, Customers, Leads, Documents, Websites, Payment / Subscriptions, MissedCallRetries, Analytics, Usage, and more). Persistence uses Prisma on PostgreSQL (including vector-backed document chunks for RAG). Background work runs on BullMQ and Redis — website crawls, campaign dial shards, missed-call retries, booking reminders, and Stripe usage reporting.

The voice path is the core differentiator: Twilio webhooks and Media Streams open a live session; OpenAI Realtime drives conversation and tool calls (calendar, knowledge search, SMS, handoff); ElevenLabs streams telephony-grade TTS back into the call. Recordings land in AWS S3; transcripts and call outcomes feed CRM and analytics.

Integrations are first-class Nest modules with OAuth and webhooks: Meta Lead Ads + Instagram / Messenger, HubSpot, Square, WhatConverts, and Google Calendar. Billing is Stripe Checkout + webhooks with metered voice usage. Observability uses Sentry; production deploys run on AWS EC2 behind Nginx with PM2 process management for API and frontend.

This architecture supports the full lead lifecycle: capture (ads, forms, social, inbound call) → qualify / converse (AI voice or chatbot) → book (Google Calendar) → persist (CRM + transcripts) → follow up (campaigns, SMS, retries).