Flagship case study2024 — PresentLive

Production AI automation, not an isolated chatbot demo.

Anita24 is one example of how I turn a messy business workflow into a production system: WhatsApp automation, realtime voice calls, RAG, agent workflows, operator handoff, deployment and monitoring working together.

System summary

Role: AI Systems Consultant · Full-stack & LLM engineer
Channels: WhatsApp + voice
Core stack: Next.js · Python · Node.js
AI layer: RAG · agents · realtime

Discuss systems work

Positioning

The value is designing the whole system around the business workflow.

The senior work is turning a complex operational need into architecture that can run in production: clear boundaries, reliable state, AI orchestration, integrations, observability and a user experience people can operate.

Business workflow to architecture

I translate unclear operational needs into product architecture: data models, workflows, AI orchestration, integrations and deployment paths.

AI connected to real operations

The AI layer connects to retrieval, tools, business rules, human review and observable production behavior instead of stopping at a chat UI.

End-to-end ownership

I work across frontend, backend, realtime channels, cloud infrastructure, monitoring and product UX so the solution is usable and maintainable.

Examples

The case study is made of real operating systems, not presentation screens.

Each system required architecture, implementation, production tradeoffs and operational thinking.

WhatsApp AI automation system

A messaging flow with conversation state, streaming AI responses, provider retries, tool calls and operator handoff.

Realtime voice and phone agent

A low-latency voice loop for natural calls, intent routing, escalation and business workflow execution.

RAG knowledge pipeline

An ingestion and retrieval system with chunking, embeddings, retrievers, reranking and domain-specific context assembly.

Agent workflow orchestration

Routing, tool selection, guardrails, human-in-the-loop review and monitoring around live customer interactions.

Engineering challenge

Make WhatsApp, phone calls and business knowledge behave like one reliable AI system.

The business needed AI automation that could support real conversations across WhatsApp, phone and chat without losing context, reliability or operator control.

WhatsApp and chat flows needed streaming answers, durable state, provider retries and operator handoff.

Voice calls needed natural latency while still supporting routing, tool calls and escalation.

Knowledge workflows needed ingestion, chunking, embeddings, retrieval, reranking and evaluation.

Architecture

Adapters, RAG and agent orchestration were built around shared conversation state.

The system was organized around channel adapters, normalized events, retrieval context, tool execution, persistence, monitoring and escalation.

Layer 1

Channel adapters

WhatsApp, realtime voice, chat and API events enter through isolated provider adapters.

WhatsAppVoiceChatAPI

Layer 2

Conversation model

Events are normalized into state, history and retryable domain events for every workflow.

Shared stateHistoryRetries

Layer 3

RAG pipeline

Ingestion, chunking, embeddings, retrievers and reranking provide domain-specific context.

IngestionChunkingEmbeddingsReranking

Layer 4

Agent workflows

Routing, tool calling, guardrails and human-in-the-loop review decide the next action.

RoutingTool callingGuardrailsHITL

Layer 5

Operations layer

Persistence, analytics, deployment and monitoring close the loop for production quality.

PostgreSQLRedisDeployMonitoring

Reliability loop

Provider errors, tool timeouts and partial failures feed retries, fallbacks and human escalation paths instead of disappearing inside a model call.

Cost and quality loop

Conversation outcomes, retrieval tests and analytics tune chunks, retrievers, tools and guardrails before increasing model usage.

Step 1

Channel adapter

WhatsApp, voice, chat or API events enter through provider-specific adapters.

Step 2

Normalize

Provider payloads become shared conversation state, history and retryable events.

Step 3

Retrieve

Ingestion, chunks, embeddings, retrieval and reranking prepare domain context.

Step 4

Agent action

The agent selects tools, routes the request or streams the next response.

Step 5

Review & observe

State, analytics, monitoring and human handoff close the production loop.

System design

The important work was everything around the model.

The core work was connecting LLMs to realtime channels, RAG context, business tools, operator controls, deployment paths and reliability loops.

LLM and agent orchestration

Separates conversation state, routing, tool calls, escalation rules and model interaction so workflows can evolve independently.

WhatsApp and realtime voice

Connects low-latency phone conversations with async WhatsApp and chat flows without duplicating provider-specific logic.

RAG pipelines

Covers ingestion, chunking, embeddings, retrieval, reranking and context assembly for domain-specific answers.

Production reliability

Designs for provider retries, partial failure, tool timeouts, monitoring and human escalation instead of assuming the model always succeeds.

Tradeoffs

Production AI required explicit choices about speed, quality and control.

The architecture was shaped around the constraints that appear when AI agents are connected to live WhatsApp and voice interactions.

Realtime latency vs. workflow depth

Keep the voice and WhatsApp response loop fast, then call deeper tools only when intent, context and confidence justify it.

Autonomous agents vs. human review

Let agents handle routine routing and answers, but keep human-in-the-loop paths for low-confidence, sensitive or failed workflows.

Retrieval quality vs. cost control

Improve ingestion, chunking, retrieval and reranking before solving quality problems by increasing model usage.

Provider payloads vs. shared state

Normalize WhatsApp, voice and chat events into one conversation model so retries, analytics and handoff stay consistent.

Engineering challenges

Where the work went beyond prompts and UI screens.

Keeping realtime voice conversations natural while still invoking tools, routing logic and business workflows

Maintaining reliable WhatsApp and chat state across streaming responses, retries, provider payloads and operator handoff

Improving answer quality with retrieval, reranking, evaluation, guardrails and cost-aware LLM orchestration

Impact

Concrete production areas delivered for Anita24.

The proof is the operating surface: WhatsApp automation, realtime voice, RAG pipelines, agent workflows and handoff paths.

AI flows

Messaging automation with state, streaming responses, tool calls and escalation paths

Voice

Realtime

Low-latency phone automation with routing, escalation and workflow execution

RAG

Pipeline

Knowledge ingestion, chunking, embeddings, retrieval, reranking and domain context

Agents

Tools

Routing, tool calling, human-in-the-loop review and workflow automation