Real systems, dated, verifiable.

A French-speaking advisory firm specialized in company setup in the Middle East needed to turn each of its case studies into three formats — a LinkedIn post, an Instagram carousel, and an avatar video (Reel) — in a strict brand identity. Delivered: a two-tier agent pipeline that picks the next case study, decides the formats, writes in the brand voice, produces a full AI video (synthetic voice + realistic avatar), has every piece approved by a human on Telegram, publishes, then answers comments automatically.

A Shopify store of around 96 products depended on one person for its day-to-day operations. Delivered: an AI agent on a VPS — 21 skills, 4 weekly cron jobs, local semantic memory, three layers of guardrails.

A generative, AI-driven engine that turns a single market theme into a complete sales infrastructure — offers, pages, creatives, funnels — deployed online and genuinely transactional. A multi-agent chain feeds a SQL database at every step; deliverables are grounded on a knowledge base (homemade RAG) and validated by schema; the Stripe payment was verified with an end-to-end test purchase.

Evening counts, critical thresholds, a shopping list passed along by hand: a process that depended on one person's memory. Delivered: a web app (Next.js, Supabase) in production on Vercel plus an iOS/Android mobile app (Expo) on the same backend.

Two complete systems — public site, back office, secured API, voice agent, invoicing, automated emails — built on the same architecture, live and verifiable today. Presented as complete projects, not client references.

A lab project, no client: read trading signals posted on Telegram, verify them, and execute orders on MetaTrader 5 without real-time human supervision. A filtered Telegram relay, a sub-second regex parser with LLM fallback, three blocking pre-trade checks, nightly ML retraining under hard limits.