---

  whoami

name:        Maxim
alias:       mmgpro
role:        AI / Automation Engineer
focus:       LLM agents · Telegram bots · Content tooling · Workflow automation
mode:        > freelance projects · long-term contracts · consulting
languages:   Python · TypeScript · Bash
ships:       fast, documented, and actually working

I build the boring stuff so you don't have to. Custom AI agents, scrapers, parsers, Telegram bots, content pipelines, internal tools. Anything that should run by itself — I make it run by itself.

---

  services.list()

> for_clients/ 🤖 AI agents on Claude / GPT / open-source models 📨 Telegram bots — from quick MVP to production 🧲 Parsers & scrapers — TG, web, marketplaces ⚙️ n8n / Make automations — wire your tools together 📊 Content pipelines — generate, post, repurpose 🛠️ Custom internal tools — dashboards, CRM-lite, ops

> for_employers/ ✅ Ship-first mindset, no astronaut architecture ✅ Comfortable with the full stack — back, front, infra ✅ Production AI agents that survive contact with reality ✅ Async-friendly, documented PRs, honest standups ✅ Russian (native) · English (working proficiency) ✅ Available remote · part-time · full-time

---

  arsenal.2026/

Top of the AI / automation stack right now — what I actually ship production with in 2026.

_{LLM CORE · AGENTS · MCP}

_{RAG · VECTORS · LONG-TERM MEMORY}

_{VOICE · MULTIMODAL · MEDIA}

_{SELF-HOSTED LLM · EVAL · OBSERVABILITY}

_{LANGUAGES · FRONTEND · DATA · INFRA}

---

  agents.deployed()

Production AI agents I've built and operate. Not demos — they run 24/7.

🤖 JARVIS  ·  flagship case study Always-on autonomous agent running on my own Linux box. Wakes on Telegram messages from a single owner-verified channel, routes tasks to specialist sub-agents (dev / devops / content / security), holds a 4-layer memory that survives across sessions, and is hardened against prompt-injection in the wild. Stack:   Claude Opus · Python · SQLite · Telegram Bot API · systemd Skills:   persistent memory · agent routing · security audit · code review · content ops Uptime: 24/7  ·  Sessions: continuous  ·  Access: owner-only

agent: Jarvis runtime: Linux + systemd brain: Claude Opus 4.7 memory: ├─ semantic (facts) ├─ episodic (events) ├─ verbatim (chunks) └─ reflection (patterns) guards: ├─ owner chat_id verify ├─ prompt-injection log └─ secret redaction

> other agents in the field/

Agent	Job	Channel	Status
🧾 Terminator	summarises group chats · 24h digests · multi-chat	Telegram groups	live
📡 Claudebot	owner-only AI controller for the home server	Telegram DM	live
🎬 Content pipeline	auto-script · SEO · thumbnail prompts	internal	live
🔒 Client agents	custom AI workers · under NDA	private	references on request

_{If you need an agent that actually does the job instead of demoing — that's what I build.}

---

  rag.expertise()

Retrieval-Augmented Generation done right. Hybrid search, two-stage reranking, semantic chunking — the boring details that keep LLMs from hallucinating in production.

> what I deliver 📚 Knowledge bases over docs · PDFs · codebases · chat history 🔍 Hybrid retrieval — BM25 + dense vectors + metadata filters 🎯 Two-stage retrieval with cross-encoder rerankers 🧩 Smart chunking — semantic · parent-child · contextual (Anthropic-style) 📊 RAG eval harness — Ragas + custom golden sets, regression-tested 🧠 Long-term agent memory — semantic · episodic · verbatim layers 🔄 Incremental indexing with dedup, versioning & TTL 🛡 Citation tracing + answer-grounding checks (no silent hallucinations) ⚡ Streaming RAG with prompt-cache warm-ups for sub-second TTFT

vector_db: Qdrant · pgvector · Weaviate · Pinecone · Chroma embeddings: OpenAI text-embed-3 · Voyage-3 · Cohere v3 · BGE-M3 (multilingual) rerank: Cohere Rerank 3.5 · Voyage Rerank · Jina parsing: Unstructured · LlamaParse · Docling · pdfplumber frameworks: LlamaIndex · LangChain · Haystack · DSPy eval: Ragas · LangSmith · Braintrust · custom memory: Mem0 · Letta · Zep · custom 4-layer context: Anthropic Contextual Retrieval (49% recall ↑)

_{Need a RAG that answers from your data and cites sources — not a chatbot that makes things up? DM me.}

---

  now.building()

Two private commercial projects currently shipping. Code under NDA — architecture and stack are open.

🎵   Reaction Engine music-news automation platform End-to-end pipeline that turns RSS feeds, Instagram / TikTok accounts and webhooks into ready-to-publish Telegram posts. AI filters relevance, kills duplicates on three levels (URL · title · semantic), generates copy + images, schedules in timezone-aware queues and ships digests from history. stack: TypeScript · React 19 · Vite backend: Express · Supabase · Caddy ai: OpenRouter · Gemini · Firecrawl ingest: Apify · RSS · Webhooks deploy: Docker · VPS status: delivered & running scope: full-stack · admin UI · client docs

🎬   Content Factory · SaaS multi-tool content platform for creators Telegram Mini App where users turn product links into UGC reels, talking AI avatars, auto-edited videos with subtitles & color, carousels, scripts, voiceovers and competitor analytics. Credit-based billing, render farm on Hetzner, tariffs from Free to Agency. stack: Next.js 16 · React 19 · Tailwind 4 data: Supabase · Redis · BullMQ workers: Python · FFmpeg · Whisper · VAD render: Remotion · Go ffmpeg-over-ip ai: OpenRouter · Gemini Veo deploy: Vercel + Hetzner VPS status: active development scope: full-stack SaaS · billing · workers

_{Need a system like one of these for your business? DM me — I take projects of any scope.}

---

  stats --dump

---

  contribution.snake

---

  let's build something

Looking for an engineer who ships? Got an idea that needs to become a product?
I'm one message away.

_{more channels arriving soon}