Marcin RapaczSenior Frontend Developer / AI Engineer

An app for people writing books — an editor with version history, a chat, style analysis, PDF import. It is not a generator that writes the book for the author: it supports work on the author's own text — ask a question, consult an idea, look for inspiration — and the system knows the book and answers from its content. I've run it from the first commit through to keeping it alive in production.

AI jobs go through a queue in the database, so a failure on the model's side never blocks saving text, and the credit returns to the user if the answer never starts. The style analysis deliberately uses no AI: it's an ordinary algorithm, so it costs nothing and gives the same answer every time. An existing book is imported by a separate agent (LangGraph) that detects chapters in code, and the model only gets a say when there are several conventions; instead of a progress bar the author reads what the agent is doing right now, and a failed import refunds the credits.

The second version of a B2B contact-sourcing tool for a client in HR, rewritten from scratch; the first was an agent with an LLM gate before sending, the second shifts the weight onto verifiable facts. Instead of buying contacts, the system builds its own base: it collects job postings with a history of bumps and paid promotions, pulls employer profiles, and a model with search access fills in what the posting lacks — company size, membership in a large organisation, phone, website. The model speaks once, in a typed answer that forbids guessing; whether a company is in the target is then decided by a plain rule on the facts.

No fact from the model reaches a person without confirmation from an independent source: first the company's website and the state register, because they are free, and only then paid Google Places with a double identity check. Every fact carries its origin and status, every model call has its cost recorded, and provider limits are enforced in code. At the end there is a list of companies with a confirmed number — a person makes the call. Over 300 offline tests, including a fake model.

Company website of a family-run foil packaging manufacturer and sweets wholesaler near Kraków, rebuilt from scratch on Next.js 16 and React 19. I ran the commission directly with the client, no intermediaries: scope, copy and SEO, launch and maintenance — everything a small company needs from one person. Solutions sized to the client: content in code instead of the previous version's headless CMS, because with a dozen or so products that means fewer services to maintain and no dependency on an external vendor.

A product photo flies from its list card into the details window and a category photo from the home page into the landing hero (View Transitions); with reduced motion enabled the transitions disappear. Analytics and the Google map load only after consent, search engines see the business as LocalBusiness, and the deployment is a Docker image behind Traefik with a GitLab CI pipeline.