// the find
neural-maze/ava-whatsapp-agent-course
Meet Ava, the WhatsApp Agent
A structured course teaching you to build a multimodal WhatsApp AI agent using LangGraph, Qdrant, Groq, ElevenLabs, and image generation via FLUX. It covers the full stack from agent graph design to Cloud Run deployment, with both written lessons and video walkthroughs. Aimed at engineers who want a real end-to-end project rather than toy examples.
The course structure is genuinely end-to-end — you get STT, TTS, image input/output, long-term memory, and WhatsApp integration in one cohesive project rather than isolated tutorials. Using LangGraph for the agent workflow is a reasonable production choice; the graph/nodes/edges structure in the source is clean and navigable. The free-tier stack (Groq + Qdrant Cloud + Together AI) means you can finish the course without a credit card. Deployment via Cloud Run with a provided Dockerfile and cloudbuild.yaml is a real deployment path, not hand-waving.
The 'long-term memory' via Qdrant is a vector store lookup over conversation snippets — it will hallucinate false memories or miss context the same way every naive RAG system does, and there's no evaluation of retrieval quality anywhere in the repo. The WhatsApp integration depends on Meta's Cloud API, which requires business verification and can be revoked; there's no fallback and the docs don't flag this risk. The schedule/proactive messaging feature (Ava sends unprompted updates about 'daily activities') is powered by a cron-triggered prompt with hardcoded fictional context — it's a demo trick, not a real autonomous agent. This is also nearly a year old and the AI tooling it depends on (LangGraph, Groq model IDs, ElevenLabs API) has had breaking changes since October 2025.