Telegram Bot Development: Command, Mini App, Conversational

Choosing the wrong Telegram bot architecture often leads to weeks of rework and thousands of dollars in additional costs. We have seen clients who started with a simple command bot for a complex multi-step order flow, only to hit the 64-byte callback_data limit and have to rewrite everything as a Mini App. Our engineers have built over 30 automated assistants in the last 5 years — from simple broadcast bots to full-blown marketplaces. That experience lets us pick the correct architecture from the start, saving your budget and timeline. Our approach reduces development time by 40% compared to starting from scratch.

Choosing the right bot type: command, Mini App, or conversational

Start with the user scenario. A simple command bot with menus and buttons is fine for order status inquiries or FAQ. You can implement it with InlineKeyboardMarkup and FSM on aiogram or python-telegram-bot. A conversational bot with NLP requires integration with Dialogflow, Rasa, or GPT API for response generation. A full Mini App is a web application (React, Vue, or plain HTML/CSS) inside Telegram's WebView, giving you a custom interface and complex business logic.

We develop all three types. Most of our clients choose a Mini App with Telegram auth — it provides seamless UX without login/password. According to our metrics, Mini Apps convert 5x more target actions compared to command bots for complex scenarios. Contact us to evaluate your project — we will propose the optimal solution without unnecessary costs. Command bots start at $500, conversational bots from $2,000, and Mini Apps from $5,000.

Why do webhooks beat polling for production?

Telegram Bot API offers two update methods: getUpdates (long polling) and webhooks. Long polling is convenient for local development — just use ngrok to test. But in production, it fills the queue: after a restart, your bot will process thousands of pending updates, potentially blocking the server. A webhook works event-driven: Telegram sends a POST to your HTTPS endpoint, and the bot responds instantly. You save bandwidth and CPU. Our benchmarks show webhooks have 3x lower latency than polling under high load and reduce server load by 80%.

Mandatory requirements for webhook:

• Valid TLS certificate (Let's Encrypt or paid)
• Port 443, 80, 88, 8443 (non-standard ports are not supported)
• Unique URL (often use /{BOT_TOKEN} for isolation)

Example webhook registration with aiogram 3.x:

from aiogram import Bot, Dispatcher
from aiogram.webhook.aiohttp_server import SimpleRequestHandler

bot = Bot(token=BOT_TOKEN)
dp = Dispatcher()

async def on_startup():
    await bot.set_webhook(
        url="<YOUR_PUBLIC_URL>",
        drop_pending_updates=True
    )

The parameter drop_pending_updates=True is critical on restart: without it, the bot will process all queued messages, causing a flood.

How to overcome the 64-byte callback_data limit?

Each InlineKeyboardButton has a callback_data limit of 64 bytes. For simple yes/no actions, it is enough. But for complex dialogs (like multi-step flight booking), you will run out of space.

The solution is to store state in Redis with a short identifier and a 1-hour TTL for automatic cleanup:

import uuid, redis, json

r = redis.Redis()

async def create_callback(data: dict) -> str:
    callback_id = str(uuid.uuid4())[:8]
    r.setex(f"cb:{callback_id}", 3600, json.dumps(data))
    return callback_id

async def resolve_callback(callback_id: str) -> dict | None:
    raw = r.get(f"cb:{callback_id}")
    return json.loads(raw) if raw else None

This approach bypasses the limit and provides a convenient FSM state. Redis automatically cleans expired entries via TTL. This technique is battle-tested on dozens of projects — we guarantee stability and performance.

Verifying initData in Telegram Mini Apps

When your bot runs as a Mini App, window.Telegram.WebApp passes initData — a string with an HMAC signature. The server must validate the signature before trusting user data (telegram_id, username). Otherwise, anyone can forge a request on behalf of another user.

Verification algorithm:

1. Derive a secret key: HMAC-SHA256 of WebAppData (as message) and bot_token (as key).
2. Parse initData by &, sort parameters (excluding hash) by key.
3. Concatenate them into data_check_string using \n as key=value.
4. Compute HMAC-SHA256 of data_check_string with the secret key.
5. Compare the result with the hash from initData.

import hmac, hashlib

def verify_telegram_init_data(init_data: str, bot_token: str) -> bool:
    secret_key = hmac.new(b"WebAppData", bot_token.encode(), hashlib.sha256).digest()
    params = parse_qs(init_data)
    data_check_string = "\n".join(
        f"{k}={v}" for k, v in sorted(params.items()) if k != "hash"
    )
    computed_hash = hmac.new(secret_key, data_check_string.encode(), hashlib.sha256).hexdigest()
    provided_hash = params.get("hash", [""])[0]
    return hmac.compare_digest(computed_hash, provided_hash)

The algorithm is based on HMAC. Verification takes less than 1 millisecond. After verification, you can use telegram_id to link with your CRM, for example, to associate with an order or subscription. Contact us for design and debugging — we guarantee correct implementation.

What's included in the work: deliverables

We offer a full cycle from idea to deployment. The scope includes:

• Scenario analysis and prototyping
• Stack selection: aiogram, grammy, react-telegram-web-app
• Backend implementation: API, Redis, PostgreSQL (or MongoDB)
• Mini App design adapted to Telegram theme (light/dark)
• Payment integration via Telegram Stars (or external payment gateways)
• Webhook setup with monitoring (uptime, alerts)
• Testing: unit tests, e2e tests against Telegram Bot API
• Debugging and publishing: submission for verification, working with App Store Review Guidelines

Deliverables after completion:

• Full API documentation
• Server credentials and access
• One week of post-launch support
• Monitoring dashboard

We have been building bots of varying complexity for over 5 years. In that time, we have launched over 30 projects — from simple newsletters to marketplaces with hundreds of products via Telegram Mini App. Every project comes with a quality guarantee and technical support. Our packages start at $500 for command bots, $2,000 for conversational, and $5,000 for Mini Apps.

Get started with your bot project

If you have an idea, write to us. We will assess the project, propose an architecture, and give timelines. Expect the following ranges:

• Command bot: from 1 to 2 weeks
• Conversational bot with NLP: from 2 to 4 weeks
• Mini App: from 4 to 8 weeks

Cost is calculated individually, depending on scenario complexity and integrations. Contact us, and we will prepare a technical specification and an accurate estimate. Get a free consultation. Our clients save an average of 30% by choosing the right architecture upfront.

Machine Learning in Mobile Apps: CoreML, TFLite, and On-Device Models

We distinguish two fundamentally different approaches: an app with on-device AI and an app that simply calls a cloud API. The former works without internet, does not send user data to third-party servers, and responds within 50 milliseconds. The latter depends on network latency and pricing plans. Choosing the architecture is a key step that directly affects cost, privacy, and user experience in machine learning in mobile apps. Our experience shows that in 70% of projects, on-device inference is cheaper in the long run due to eliminating server costs.

How to Choose Between CoreML and TFLite for On-Device Inference?

CoreML — Apple's native framework for running ML models on device. Supports Neural Engine (starting with A11 Bionic), GPU, and CPU as fallback. Models are converted to .mlmodel format via coremltools from PyTorch, ONNX, or TensorFlow. Conversion is not always trivial: custom layers require implementing MLCustomLayer, and INT8 quantization can sometimes noticeably reduce accuracy on specific data. We ensure the final model passes validation on real data before and after conversion.

TensorFlow Lite — cross-platform alternative for Android and Flutter. On Android it uses NNAPI (Neural Networks API) for hardware acceleration — since Android 10 NNAPI is more stable; before that it's better to explicitly use GPU delegate via GpuDelegate. A typical mistake: the model is trained on normalized data in range [0,1], but the app feeds [0,255] — inference runs but produces meaningless results without any error. We include an automatic input data validation module in the SDK.

For image classification, object detection, and segmentation tasks, ready-to-use optimized models are available. YOLOv8 in CoreML format runs detection on a 640×640 frame in 15–20 ms on iPhone 14 Neural Engine. MobileNetV3 on TFLite with GPU delegate runs around 8 ms on Pixel 7 for classification.

What If You Need Text Generation On-Device?

Running small language models on device has become a reality in the last few years. Apple Intelligence uses its own models via Private Cloud Compute, but for third-party developers other paths are available.

llama.cpp with Metal backend on iOS is a working approach for phi-3-mini (3.8B parameters, 4-bit quantization, ~2.3 GB). Inference: 15–25 tokens/second on iPhone 15 Pro. For integration in Swift, use the Swift Package llama.swift or a wrapper via C interface llama.h. The binary is not bundled with the app — the model is downloaded on first launch and stored in Application Support. Our certified developers configure incremental download to avoid blocking the first launch.

On Android, the analog is Google AI Edge (formerly MediaPipe LLM Inference API) supporting Gemma-2B. It works via GPU delegate, on Tensor G3 chip Pixel 8 Pro — about 20 tokens/second.

Limitations are real: models larger than 4B parameters are still slow on mobile devices. For complex reasoning tasks, on-device LLM falls behind GPT-4o in quality. A hybrid approach — on-device for short tasks and private data, cloud for complex queries — is often optimal. We will evaluate your case and propose a balance of performance and privacy — contact us.

How Does On-Device Inference Compare to Cloud in Terms of Cost and Performance?

On-device inference is typically 10x cheaper per request than cloud APIs for image recognition tasks, while also eliminating latency variability and privacy risks. The table below summarizes the trade-offs.

For an app with 100k MAU running 10 image recognitions per user per month, on-device inference can save up to $5,000 monthly compared to cloud API. Get a free consultation on your ML architecture today.

Integrating OpenAI API and Other Cloud Models

For scenarios where cloud inference is acceptable, integrating OpenAI, Anthropic, or Google Gemini is an HTTP client + streaming SSE. In Swift, AsyncThrowingStream is convenient for streaming responses. In Kotlin, use Flow.

Critically: API keys must never be stored in the app bundle. Even an obfuscated key can be extracted from the IPA in 10 minutes using strings or frida. Correct architecture: mobile app → your own backend → OpenAI API. The backend controls rate limiting, logs requests, and protects the key.

What Is Included in the Work (Deliverables)

• Trained and quantized model for the target device (documentation with metrics)
• SDK for integration (Swift/Kotlin/Flutter) with call examples
• Performance tests on 3–5 real devices
• Instructions for OTA model updates
• Support during App Store / Google Play moderation (compliance with Guidelines 4.2, 5.1)
• 2 weeks of technical support after release

Typical Project Pipeline

1. Task analysis — measure latency, privacy, size, supported devices.
2. Model prototyping — in Python, evaluate accuracy on target data.
3. Conversion and quantization — for CoreML/TFLite with validation.
4. Integration into the app — model wrapped in a service layer (easy to swap CoreML ↔ TFLite ↔ cloud).
5. Testing — on real devices, measure FPS, RAM, battery.
6. Deployment — via TestFlight / Firebase App Distribution, monitor metrics.

Timelines: integration of a ready CoreML/TFLite model — 1–2 weeks, development of a custom model with mobile optimization — from 6 weeks, on-device LLM chat with personalization — 4–8 weeks.

Why We Take on Complex Cases?

10+ years of experience in mobile development, 50+ implemented AI/ML solutions, guarantee of compatibility with current iOS and Android versions. All projects undergo code review and load testing. The cost includes preparation of moderation documentation and training of your team.

Contact us — we will help you choose the architecture and implement ML in your app turnkey. Order an audit of your existing solution — we will assess the potential for server cost savings free of charge. In some projects, savings can reach significant amounts per month.