I turn business needs into working software, and lead the teams that build it.

Healthcare · Public hospital

Multilingual voicebot for patient appointments

A voice agent that calls patients, answers their questions and hands over to a person when needed.

• Azure OpenAI Realtime
• Azure Communication Services
• TypeScript
• Nuxt
• PostgreSQL + pgvector
• Bicep

Problem

Staff made a huge number of calls to remind patients of appointments, confirm them and explain how to prepare for tests.

Built

A real-time voice agent that calls patients in several languages, answers questions with RAG, hands over to a person when needed and shows live calls on a web dashboard. Built with infrastructure as code and 600+ automated tests.

My role

Solution architecture and technical lead.

Insurance · International group

Automated audits of websites worldwide

A platform that analyses websites worldwide and scores them against brand, accessibility and security guidelines.

• Azure Functions
• Service Bus
• Bicep
• Playwright
• React
• Azure OpenAI

Problem

Thousands of websites had to be checked by hand against brand, accessibility and security guidelines.

Built

An event-driven serverless platform that analyses each site and scores it by combining deterministic image analysis and LLMs. Backed by 250+ automated tests.

My role

Solution architecture and delivery lead.

Own product · SaaS

Multi-tenant platform for building AI assistants

A shared platform where each organisation builds and publishes its own AI assistants, with isolated data hosted in the EU.

• Azure Container Apps
• Azure OpenAI
• PostgreSQL row-level security
• Entra ID
• Terraform
• RAG

Problem

Each new client needed its own full cloud installation, with fixed costs and no way to sell the product as a service.

Built

A shared platform where each organisation builds, trains and publishes its own assistants, with data isolated per tenant, everything hosted in the EU, usage-based cost control and a dedicated mode for clients who need it in their own cloud.

My role

Architecture, technical roadmap and business case: costs, pricing and go-to-market plan.

Public sector · Regional administration

AI review of complete administrative case files

Multimodal AI that reads the whole case file, checks it against the procedure requirements and drafts a proposed decision.

• Multimodal LLM
• Amazon Bedrock
• Textract
• AWS Lambda
• ECS
• Terraform

Problem

Staff checked by hand that every application file was complete, correct and met the requirements before proposing a resolution.

Built

A multimodal AI pipeline that reads the whole file (scans, forms and attachments), classifies every document, checks it against the requirements of the procedure and drafts a proposed resolution for the case officer. It works for any administrative procedure. Deployed with infrastructure as code.

My role

Solution architecture and technical lead.

Own product · SaaS

Automated Google review management for businesses

A subscription service that replies to each business’s Google reviews in its own voice, with one-click approval from email.

• Next.js
• TypeScript
• PostgreSQL
• OpenAI
• Google Business Profile API
• Stripe
• Coolify

Problem

Small businesses have no time to answer their Google reviews, and an unanswered or badly answered review costs them customers.

Solution

A subscription service that syncs reviews from Google Business Profile, classifies them with AI and drafts each reply in the tone and with the knowledge of each business. Replies go out automatically or after one-click approval from an email. It includes QR codes to ask for reviews, monthly PDF reports, payments and a log of every AI decision.

Architecture

A single Next.js application (customer dashboard, admin and public routes) with PostgreSQL and scheduled jobs, deployed with Coolify on a Hetzner server. Built with a Claude Code agent harness: leader, implementer and reviewer.

Personal finance · AI

Personal wealth tracker with an AI assistant connected to the banks

All accounts and investments in one place, with an AI assistant that answers with charts.

• Python
• FastAPI
• SQLite
• Open banking (PSD2)
• Claude API
• MCP + OAuth
• Docker
• Cloudflare Tunnel
• Hetzner

Problem

Money was spread across banks, a broker and an index fund, with no single place to see it all or ask questions about it.

Solution

A web app that connects to the banks through open banking (PSD2) and to the investment manager’s API, categorises transactions, detects transfers and builds forecasts and projections. It has an AI assistant that queries the data with tools and answers with charts, push notifications on the phone and an MCP server with OAuth so Claude can access it.

Architecture

FastAPI and SQLite in Docker containers on a Hetzner server with no open ports: it is reached through a Cloudflare tunnel protected by Cloudflare Access. Automatic deployment with GitHub Actions.

Project management · MCP

Project and team capacity control

Budget, hours and team capacity in a single app, with alerts and an MCP server to query it from Claude or ChatGPT.

• Next.js
• React
• TypeScript
• MCP
• Playwright

Problem

Budgets, hours and team availability were tracked in a spreadsheet that gave no warning of overruns or over-allocation.

Solution

An application for the project portfolio: budgets and extensions, actual and forecast hours per person and month, rate periods, capacity with public holidays, leave and other commitments, and alerts. It reconciles hours with the corporate timesheet portal and imports leave from it. Its MCP server lets Claude or ChatGPT read and change the data under the same rules as the interface.

Architecture

Next.js and React with the API in the same process. The timesheet portal connector runs on Playwright and stores credentials encrypted by the operating system.

Sport · AI

AI running coach connected to a Garmin watch

An AI coach that adapts sessions to sleep and recovery, schedules them on the watch and talks back by voice.

• Python
• FastAPI
• SQLite
• Garmin Connect
• Connect IQ (Monkey C)
• Claude API
• OpenAI speech-to-text
• MCP + OAuth
• Docker

Problem

Generic training plans ignore how you slept, how you have recovered and what you have actually run, and changing workouts by hand on the watch is slow.

Solution

A web app and a watch app that sync activities, sleep, HRV and readiness from Garmin Connect and compare planned against completed training. A tool-using AI coach creates and reschedules workouts and plans, and only schedules them on the watch once confirmed, with history and undo. On the watch you talk to it and it answers by voice. It has push notifications and an MCP server with OAuth so it can be used from Claude.

Architecture

FastAPI and SQLite in Docker containers on a Hetzner server, behind a Cloudflare tunnel protected with Cloudflare Access. The watch app is built with Connect IQ (Monkey C). Automatic deployment with GitHub Actions.