Skip to main content
Back to projects

Case Study

OMNICHANNEL AI ASSISTANT PLATFORM

Role

Backend & AI Engineer

Platform

FastAPI Automation & RAG Service

Core capabilities

  • FastAPI
  • RAG Pipeline
  • Pinecone Vector DB
  • WhatsApp API
AI Assistant Platform — application interface

What is AI Assistant Platform?

An omnichannel AI customer-engagement and workflow automation platform built as a personal R&D system exploring Python and applied AI architecture. It pairs a React operational dashboard with a FastAPI intelligence service—combining Pinecone vector search (RAG), document ingestion, STT/TTS voice pipelines, and Meta WhatsApp Cloud API webhooks for automated customer support.

What I built

  • Hybrid Laravel + FastAPI Compute Architecture. Separated operational workspace management, authentication, and webhook routing from AI compute by pairing a React 19 dashboard and administrative API with an asynchronous Python/FastAPI intelligence service.
  • Workspace-Scoped Vector RAG Pipeline. Built an end-to-end document ingestion pipeline that parses uploaded knowledge-base files, splits text into overlapping semantic chunks, generates vector embeddings, and executes metadata-filtered similarity queries against Pinecone scoped per workspace.
  • Low-Latency Voice AI Pipeline (STT → RAG → TTS). Engineered a voice interaction pipeline chaining OpenAI Whisper speech-to-text transcription, context-injected LLM response generation grounded in retrieved vector chunks, and streaming text-to-speech synthesis for automated voice support.
  • Omnichannel Webhook Routing & Guardrails. Integrated Meta WhatsApp Cloud API, Instagram, and Facebook Messenger webhooks with signature verification, conversation-state tracking, confidence thresholds for human-agent handoff, and per-workspace usage metering.

System scope

  • Decoupled MicroservicesLaravel core business engine paired with high-performance Python/FastAPI AI engine.
  • RAG Document IngestionChunking, embedding, and vector retrieval grounded in workspace-scoped Pinecone indexes.
  • Multimodal Voice PipelinesReal-time speech-to-text (Whisper) and text-to-speech (ElevenLabs) conversational engine.
  • Workspace-Scoped RBAC & Webhook RoutingStrict data partitioning, workspace invitations, audit logging, and subscription billing.
  • Observability & MonitoringTelemetry tracking token usage, vector query latency, and automated retry policies.
  • Subscription Billing & Webhook State SyncPer-workspace usage metering, subscription lifecycle handling, and inbound omnichannel webhook state tracking.

Inside the product

The operator's view opens on the numbers that decide whether the automation is working: revenue attributed, conversation ability, funnel health out of five, and average response time, over a usage timeline you can bound by day, week, month or all time.
The Brain is the retrieval layer. Uploaded documents are listed with their size and indexing state, so a file that has not been embedded yet is visible as such rather than silently answering from nothing.
Documents is where the source material is managed - file, size, type, indexing status and chunk count per row - with storage usage shown against the plan limit and a downloadable usage report.

LET'S BUILD
YOUR NEXT SYSTEM

Whether you are hiring for a backend engineering role, architecting a multi-system SaaS platform, or automating complex operational workflows — let's connect.

0/1500

or download my technical resume →