Aibolit Community — AI brand assistant

A pet-care community wanted to give its audience an always-on assistant — and a place inside Telegram where the brand could pull free users into a paid tier. I built the assistant, the mini-app, and the partner analytics surface. Brand and content owned by the client; I was the build partner.

Stack

React + Vite

Live since

2025

Surface

Live mirrors

EN + RU

Problem

The brand had a strong free-tier audience inside Telegram — owners asking quick questions about their pets — and no path from that audience into anything paid. A Telegram bot that only forwards messages to an admin doesn't scale, and a separate web app asks the user to leave the surface they're already on.

They wanted three things on one surface: a free conversational assistant good enough that users open it daily, a paid tier with features the assistant can hand off to, and a way for partner brands to see which segments engage with which topics — without building a separate analytics product.

Solution

A Telegram mini-app with a conversational AI front and centre. Owners describe their pet's situation in plain language; the assistant replies inside the same Telegram surface, with a triage hand-off when the question crosses out of "general care".

Free tier covers everyday questions. Paid tier unlocks longer sessions, saved pet profiles and partner-curated content. The partner analytics surface is the same React app under a different role — same data, scoped view — which kept the build to one codebase.

Landing hero — diagnosis surfaced as a single sentence: «Your bot is burning 28,600 ₽/mo on tokens», anchored to the line-by-line read of 1,964 actual dialogues.

Architecture

Vite-built React app served as a Telegram mini-app, talking to a small Node API that fronts the AI provider and persists chat history per Telegram user id. The bot side handles entry, deep links and webhook subscriptions — anything that needs to live outside the mini-app frame.

Tiers are role flags, not separate apps. The partner analytics view is the same client, gated by role, reading from the same event log the assistant writes to. One codebase, three audiences.

Partner financial model — Active / Leads / Paying breakdown on the left, MRR + clinic-leads + net margin on the right. Real numbers, brand-internal.

Four sections worth pausing on

The landing page is structured as a four-step diagnostic essay: hard numbers from the actual logs, why the spend overruns, what each architectural fix does, and how the result pays back. Each section below is its own deep-link inside the live page.

«Where the money goes today» — four causes of token overrun, surfaced from a line-by-line read of 1,964 real dialogues.

Five architectural fixes — memory caching, smart search, long-dialogue compression, abuse-protection — each module pays for itself.

Phase-based scope configurator — Foundation, Monetisation, Premium AI, Scale — pricing visible per phase, no hidden bundles.

Payback model — slider for projected MAU, financial model rebuilds live, payback in months falls out as a single figure.

Results

Hard engagement numbers and partner conversion data are owned by the client and not published. What I can confirm is the shipping shape: free and paid tiers both live, one partner brand onboarded, two-locale presence kept in sync.

Under NDA

MAU

Client-tracked

Under NDA

Paid conversions

Brand internal

Free + paid

Tiers shipped

Both live

Partner brand

BestDinner

Hard numbers under NDA — available on intro call.

White-label angle

The same shape works for any audience already living inside Telegram — fitness communities, finance forums, hobby publishers. Brand owns the voice and content; I drop in the assistant, the tier model and the partner analytics. A studio can resell this into a client's existing community without rebuilding the shell from scratch.

The English clone above is a working mirror, not a marketing page. Open it on Telegram and the surface behaves the way it would inside a real brand deployment.