Metra AI — Production SaaS for Telegram Content Automation

The business problem

Telegram channel owners and content teams spend enormous amounts of time creating posts manually. Standard solutions either don't fit Telegram's specific UX (Buffer, Hootsuite are built for Instagram/Twitter), or rely on raw ChatGPT output that produces generic, low-quality content requiring extensive manual editing.

The core pain: content teams spend 60–80% of their time on production instead of strategy, and quality suffers because AI-generated content typically lacks brand voice, channel-specific context, and real-time relevance.

What I built

A full SaaS platform that automates the entire Telegram content workflow:

• AI-powered post generation that maintains brand voice and channel lore
• Scheduling system with reusable weekly presets
• Real-time data integration for news-driven content
• Integrated Telegram CRM for managing leads without exposing owner account credentials
• Multi-account, multi-operator infrastructure for agencies running multiple channels

Architecture: why multi-agent instead of single LLM call

The core technical innovation is a multi-stage LLM orchestration pipeline rather than relying on single API calls. This was a deliberate architectural choice based on a key insight:

Solution: decompose post generation into specialized stages, each with a single focused responsibility.

Standard posting pipeline

• Structure parser — extracts post anatomy and preserves links (which premium LLMs strip out by default)
• Content rewriter — handles each paragraph in isolated calls, preserving structural integrity
• Style enhancer — adds emojis and formatting based on channel persona
• Auto-rules validator — applies channel-specific rules (e.g., removing punctuation, enforcing length limits)

Extended posting pipeline (generates from scratch)

• Archetype selector — chooses post structure based on content type and length parameters
• Block-by-block generator — writes each section with focused context
• Style and formatting application
• Auto-rules validation

This architecture eliminates the typical AI failures — hallucinations mid-text, structural drift, lore over-application, format breakage — that single-shot LLM calls produce.

Key technical innovations

1. Prompt normalization layer

The image model over-refused legitimate prompts — false positives on ordinary requests like generating people. Rather than switching to a pricier model, I built a prompt-normalization layer that rephrases benign user input so it isn't incorrectly blocked, preserving the original intent. This kept quality high without hosting more expensive alternatives.

2. Lore compression and translation

User-provided channel lore (often 3000+ tokens of unstructured text) is compressed and translated to the channel's posting language at upload time. The AI receives a structured, language-matched summary instead of raw lore, dramatically improving content relevance while reducing token costs.

3. Lore as context, not constraint

Discovered through iteration that AI over-applies lore when given it as direct instruction. Architected lore as soft context that influences but doesn't dominate generation — producing more natural content.

4. Premium LLM selection strategy

Chose specific LLMs for specific reasons:

• Fewer false refusals on legitimate edge cases
• Better real-time news integration via Perplexity-style APIs
• Cost optimization at scale

Infrastructure

• Multiple Ubuntu servers in production (main service, CRM, staging)
• 16 Docker containers orchestrated with appropriate separation of concerns
• Backend as single source of truth — all client requests flow through backend, never directly to AI providers or DB
• Monitoring stack: Prometheus, Grafana, Sentry for error tracking, Uptime Kuma for service health
• Encryption: All sensitive data (phone numbers, messages, passwords) encrypted with proper salt+pepper and GPU-resistant hashing. Decryption keys held off-server.
• Security: 2FA, JWT rotation, session fingerprinting, domain proxying

Outcome

Launched to production within the 3-month development window. Multi-account CRM allows agencies to manage operators without exposing channel owner credentials. Active early-stage traction with growing user base.

Tech stack

What this demonstrates

• Ability to architect and build production SaaS solo
• Deep understanding of LLM limitations and how to architect around them
• Production-grade security and infrastructure design
• Pragmatic cost optimization at architecture level
• End-to-end product thinking: business problem → technical solution → deployment → operations