Budget: 27000 UAH Deadline: 40 days
Regarding the project. The task is clear, but we need to immediately separate the wet fantasies from the CrewAI manuals and the harsh reality of production under high load and scaling. Creating a "set of scripts" that will get banned on WhatsApp in two hours is a matter of one evening. Building a fault-tolerant ecosystem that won't crash due to API limits and will maintain the states of thousands of leads is a completely different story.
What I recommend and how it MUST be done:
ARCHITECTURE BASE. No CrewAI for production in its pure form; they are only good for demos to investors. Only LangGraph. Why? We need strict State Management, control of cycles, and determinism of the graph. Agents should not go into infinite inference and eat up your money on LLM.
WHATSAPP AND BANS. Cold sales on WA without warming up and strict policies mean death for numbers. We use a proper gateway (the same Whapi or the official Meta API, if the budget allows) + OVERLAY control. We implement Human-in-the-loop at early stages for critical actions; otherwise, the neural network will promise clients free treatment at your expense.
SCALING AND LOCALITY. If it needs to be local, we package everything in Docker Compose. Local ChromaDB is okay for a start, but for proper RAG under the laws of different countries (GDPR, HIPAA), the database will need to be correctly sectioned by metadata so that the context of Italy does not mix with Australia.
Regarding demos and cases. All combat systems operate in closed client environments under strict NDAs; I won't provide access to "poke around" as it's a commercial secret. However, during a call with screen sharing, I can easily show the architecture of working systems, the logic of graphs, state machines, and how data flows.
To start and evaluate, we need to clarify three points:
Is there already a database of dental clinics (what format) or are we also writing a parser from scratch for maps/catalogs?
How ready are the texts of offers and scripts that the agent should learn to sell?
What hardware is allocated for local launch (are we setting up small local networks for parsing or is all inference strictly on OpenAI/Claude via API)?
If reliable software with solid logic is needed, not a crooked hack - let's discuss in private messages, and I will send the technical specifications for approval.
