CreatorHive

Overview

CreatorHive is an AI-powered Telegram bot built for an online coaching community that helps people launch UGC (user-generated content) businesses. The bot operates under the persona "Ava" - a warm, casual community manager who engages new members in genuine conversation, qualifies them against a four-dimensional scoring model, and books free strategy calls without any manual outreach.

Under the hood, CreatorHive uses a hybrid finite-state-machine plus LLM architecture. A deterministic FSM built on aiogram 3 controls the conversation flow through eight core states and three branch states, while Claude Sonnet generates the actual dialogue. Every LLM response returns structured JSON containing the message text, qualification signal scores, detected objections, and a recommended state transition - which the FSM validates against a strict transition map before advancing. The result is a bot that sounds human but never goes off-script.

The Challenge

The Creator Hive Skool community had hundreds of members joining regularly, but no scalable way to separate high-intent prospects from casual browsers. The team was manually messaging new members one by one, trying to gauge interest through back-and-forth Telegram conversations - a process that was inconsistent, time-consuming, and impossible to sustain as the community grew.

Qualified leads were slipping through the cracks. People who were genuinely ready to invest in coaching never got a timely response, while hours were spent on members who were just browsing. The community needed a system that could engage every new member instantly, qualify them against objective criteria, handle objections gracefully, and route only high-intent prospects to the sales team - all without feeling robotic or pushy.

The Solution

We built "Ava," an AI Telegram bot that qualifies prospects through natural conversation. The core architecture separates control flow from language generation: a deterministic state machine (aiogram 3 FSM) governs conversation progression through states like Greeting, Pain Discovery, Urgency Check, Qualification Summary, Booking Offer, and Objection Handling, while Claude Sonnet generates context-aware responses within per-state prompt constraints. Each state has explicit transition rules - the LLM can suggest a transition, but the FSM validates it against an allowed-transitions map before executing.

Qualification uses a four-axis scoring model: financial readiness, pain intensity, urgency, and mindset - each scored 1 to 5. The LLM embeds these scores as structured signals in every JSON response, and prospects must hit a minimum composite score (12/20) with at least a 2 in financial readiness before receiving a booking offer. Objection handling is capped at two attempts per prospect, after which the bot gracefully exits to a follow-up state rather than becoming pushy.

The booking flow integrates with iClosed rather than Calendly, with a URL injection system that replaces any hallucinated links the LLM might generate with the real booking URL. A post-processing layer also strips em dashes and other formatting artifacts that Claude tends to produce despite prompt instructions.

The system runs on FastAPI with a PostgreSQL backend, deployed to Railway via webhook mode. FSM state is persisted to the database and automatically restored after deployments - solving the critical problem of aiogram's in-memory storage being wiped on every redeploy. Background schedulers handle automated follow-ups: 2 hours after an unbooked offer, 48 hours for "thinking about it" prospects, and 7 days for nurture sequences.

Tech Stack

Results