AI inproduction.

Sports federations have their own exercise banks and their own way of structuring training, but club coaches often plan in documents and spreadsheets. A general AI model readily suggests exercises the federation has never approved.

I founded and built Atletio, a platform where each federation has its own organisation, its own exercise bank and its own coach onboarding without code changes. Plans are generated with Claude and streamed to the coach. Before a plan is saved, the server checks that every exercise exists in the federation's bank, a rule written into the project's constitution. The login is tested so that an outsider cannot find out which email addresses belong to customers.

The agency runs some twenty local advertisers in retail, tourism and industry, with accounts in both Google Ads and Meta. Account audits, campaign builds and reports were done by hand, one account at a time.

I built MCP servers that give Claude access to Google Ads and Klaviyo, and a working method on top of them. The Google Ads server has eight read tools and ten write tools. Every write runs as a dry run until someone switches on live mode, and caps on budget, cost per click and percentage increase sit in the server configuration. Creating campaigns and ads has to be enabled separately. The Klaviyo server labels its 64 tools as read, write or destructive, so the dangerous ones can require approval.

The team working with the chain produces ads, signage and campaign material for many stores every week. Everything has to match opening hours, the week's offers and the chain's rules on pricing and legal text, and errors were caught in manual proofreading.

I built a plugin for the agency's Claude organisation with five skills and two connectors that answers questions about stores and proofreads material against real data. Around it sits a campaign portal with an MCP endpoint for store data, a sign generator where the price comes from the offers API and cannot be edited in the form, and an InDesign tool where Claude only classifies the offer type while ordinary code calculates the price. A review service checks campaign material against the product list.

The agency's clients wanted to know what is moving in social media and the news to plan content and campaigns. Monitoring that by hand takes a lot of time and is rarely complete.

I built a pipeline that every day collects data from Google Trends, Google News, TikTok, X, YouTube, Spotify and from SVT, DN, Aftonbladet and Swedish Radio. Claude analyses the material, and embeddings in a vector database group what is about the same thing. The system flags anomalies and sends reports as PDF and email. It also has its own MCP endpoint, so the reports can be read by an AI agent.

Information about living in, moving to and visiting Östersund is spread across many websites. Anyone looking has to know where to look.

I built both the chat and the back end. The back end collects content from the websites, puts it in a vector database and has an admin interface for the knowledge base. The chat answers by searching that database and shows cards for weather, places and events. A separate entry point, Move here, helps people thinking about moving to the municipality.

The agency's leadership named AI a key question for its competitiveness. Knowledge was unevenly spread, from staff who had never written a prompt to developers already working with agents.

I have developed and delivered training for staff and clients, including for the tourism industry in Jämtland on behalf of Jämtland Härjedalen Turism, with my own material and prompt library. The same knowledge sits in aikunskaper.se, a course platform where participants write prompts, get them assessed by an AI and can improve them in a chat inside the course. Organisations can onboard their staff and issue certificates.