Production RAG Sales Agent for E-commerce: da-vinchi.pl
A full-stack AI Sales Assistant for a Polish wallpaper store. LangGraph ReAct agent with 5 tools, Hybrid RAG (SQL + pgvector), ETL pipelines, buyer-scenario tests, and Telegram interface — production-grade, not a demo.
Invoice Automation Pipeline: Gmail → PDF → Gemini → Drive + Supabase
Invoices from multiple suppliers arrive as PDFs by email in different formats. Manual routine: find, classify, extract, file, enter into accounting. Now it runs on a schedule — automatically, end-to-end.
AI Image Composer: Identity Replacement in 30 Seconds
A Polish creative studio was burning hours on trial-and-error prompting for personalized AI paintings. I built a production web app in one day using React 19 and Gemini that turns it into a 30-second workflow.
From Data Mirroring to Knowledge Synthesis: RAG Architecture Evolution
My first mistake was vectorizing raw PDFs. The result was a "calculator without a brain." Here is the architectural shift — Parent Document Retrieval, LCEL chains, and two-layer storage — that actually eliminated hallucinations.
QA Automation for RAG Systems: Testing AI Pipelines
Building a RAG pipeline is one thing. Knowing it works reliably under edge cases is another. I built a testing framework with buyer-scenario integration tests, 98% data optimisation, and GitHub Actions automation.
Vibe Coding a Hybrid RAG Agent: Supabase + Gemini File Search
A 57 MB PDF catalog, 55K+ tokens, ~$45/mo in hot-cache costs. I engineered a Hybrid RAG Telegram Sales Agent combining Supabase pgvector for hard data with Gemini File Search for semantic context — sub-second responses, near-zero API cost.
Common thread across all these projects
The real engineering is not in the LLM call. It is in what surrounds it: controlled ingestion, deterministic fallbacks, idempotent pipelines, and structured storage. That combination is what separates production automation from demos.