Case Study

Visitor AI Assistant

A public-facing Telegram AI assistant built for first-contact conversations, project discovery, request drafting, moderation, and smooth handoff to the owner.

Completed successfully Python + Aiogram Session Memory Moderation + Handoff

Public-facing layer

Separate visitor contour

This project is presented as its own public product layer inside the broader Dekrov AI ecosystem. It is designed for consultations and routing, not for owner-only internals.

Built for first contact, not generic endless chat.
Keeps temporary session context while the consultation is active.
Can escalate meaningful requests to the owner without exposing private system behavior.

Problem

A public AI assistant has to do more than answer politely. If it is too generic, it wastes leads; if it is too open, it leaks internal context; if it is too rigid, it kills useful conversations. The challenge is building a first-contact layer that helps people move forward without becoming either a dumb FAQ or an unrestricted chatbot.

Solution

Visitor AI Assistant was built as a separate consultation contour: it answers public questions, helps visitors shape a meaningful request, preserves short session context, and keeps clear behavioral boundaries.

Session-based Conversation context is kept only for the active consultation and expires after inactivity.

Public-safe The assistant stays inside public-facing knowledge and avoids owner-only internals.

Owner handoff Useful requests can be turned into a cleaner message for the owner instead of forcing visitors to start cold.

Guarded dialogue It can detect drift, keep boundaries, and end low-signal conversations when the dialogue stops being useful.

Architecture

Public entry point

Visitors talk to a Telegram bot that starts and manages a consultation session instead of exposing the owner-facing userbot directly.

Consultation router

The assistant classifies requests, answers public questions, helps formulate messages, and routes useful conversations toward an owner handoff path.

Guardrails

Session state, moderation, boundary tracking, and quality review keep the visitor layer useful without letting it turn into an unrestricted generic chatbot.

Features

Pre-consultation help

Instead of just dropping links, the assistant can help visitors understand what they need, clarify a task, and prepare a message before contacting the owner.

Structured handoff

When the conversation becomes meaningful, the visitor flow can redirect the user toward a clearer owner request rather than leaving them with a vague chat log.

Short-lived session memory preserves the active consultation without growing into a heavy long-term profile system.
Supportive drafting helps users who do not know how to describe a bot, automation flow, or collaboration request yet.
Small talk is allowed in a limited form, but the assistant keeps its role as an assistant instead of pretending to be a friend.
Public answers are separated from deeper owner-only project knowledge so the visitor layer stays clean and safe.

Operations and Guardrails

Moderation layer

The flow can warn, limit, block, or end a consultation when behavior becomes abusive, repetitive, or clearly detached from the assistant's purpose.

Session lifecycle

Active consultation data is temporary by design: a session can expire after inactivity, and forced restarts do not keep unlimited conversational baggage.

The system distinguishes meaningful follow-up questions from low-signal drift or empty chatter.
Owner handoff stays explicit, so the assistant does not pretend it already completed something that still needs human review.
A separate review layer can flag poor visitor replies and repeated degradation patterns for later fixing.
The visitor contour remains separate from the owner-side assistant logic, which helps preserve safer public behavior.

Technology Stack

Python Core logic, orchestration, moderation, and session handling. Aiogram Telegram bot transport, callbacks, routing, and visitor-facing interface handling. Groq API LLM inference for consultation replies, drafting, and review logic. Asyncio Async workflows for message handling, session updates, and background checks. JSON Persistence Session state, moderation state, and lightweight runtime storage without a heavy separate backend.

Result

Visitor AI Assistant turns the first contact layer into an actual product surface: helpful enough to guide people, strict enough to keep direction, and structured enough to pass meaningful requests forward. As a portfolio case, it shows product thinking around boundaries, public UX, moderation, and AI behavior quality, not just raw chatbot generation.