Building AI Assistants: Lessons from Zyzik

At AINGERS, one of our most interesting projects has been Zyzik — an AI assistant for athletes and fitness enthusiasts. Here’s what we learned shipping an AI product from zero to production.

What Zyzik Does

Zyzik is a Telegram-based AI assistant that helps athletes with training plans, nutrition advice, and recovery tracking. It uses LLMs to understand natural language queries and provides personalized recommendations based on user data.

Lesson 1: Prompt Engineering Is 80% of the Work

We spent more time writing and iterating on prompts than writing code. A well-crafted system prompt with clear boundaries, examples, and guardrails makes the difference between a useful assistant and a hallucination machine.

What worked:

• Structured system prompts with explicit role definitions
• Few-shot examples for common query types
• Clear instructions on when to say “I don’t know”
• Temperature tuning per task type (lower for facts, higher for motivation)

Lesson 2: RAG Is Not Optional

A general LLM knows about fitness, but it doesn’t know about YOUR user’s training history. We implemented Retrieval-Augmented Generation to feed user-specific data into every response.

Our RAG pipeline:

1. User sends a message
2. We retrieve relevant context (recent workouts, goals, preferences)
3. Context + query go to the LLM
4. Response is validated and sent back

This took the assistant from “generic fitness chatbot” to “personal coach who knows your history.”

Lesson 3: Latency Matters More Than You Think

Users expect instant responses in a chat interface. An LLM taking 5 seconds to respond feels broken.

Our optimizations:

• Streaming responses (user sees text appearing in real-time)
• Caching frequent queries
• Using smaller models for simple tasks (classification, intent detection)
• Reserving powerful models (Claude, GPT-4) for complex reasoning

Lesson 4: Guard Against Hallucinations

For a fitness assistant, wrong advice can cause injury. We implemented multiple safety layers:

• Validate nutritional data against a verified database
• Flag recommendations that fall outside safe ranges
• Add disclaimers for medical-adjacent advice
• Log and review edge cases weekly

Lesson 5: Start With One Channel

We launched on Telegram only. Not a web app, not a mobile app — just Telegram. This let us iterate fast with real users, gather feedback, and improve the core AI before investing in multi-platform development.

The Tech Stack

• Python + FastAPI for the backend
• aiogram 3 for Telegram integration
• Claude API as the primary LLM
• PostgreSQL for user data
• Redis for caching and session management
• ChromaDB for vector search (RAG)

Key Takeaway

Building an AI assistant is 20% AI and 80% product engineering. The model is a component — the real challenge is building a system that’s reliable, fast, and genuinely helpful.

If you’re considering an AI assistant for your business, start small. One channel, one use case, real users. Scale what works.