Bound the assistant's role
The user assistant has no unrestricted tool access. Code controls context and output; instructions and filters bound responses. Product development and operations agents serve a different role.
Strahlkraft40+ / Digital product
Strahlkraft40+ began with a market opportunity in women's wellbeing. We defined the product with its co-founder and built a platform with an AI assistant, content, payments and operational tools.
Product co-founded by Ionela Harosa and Ignacio Aredez
Live product · case documented as of October 2026
01 / The problem and its context
Ionela Harosa, Strahlkraft40+ co-founder, identified an opportunity in women's wellbeing. She understood the audience and had a business idea, without a technical background. We worked with her to define the experience, the assistant's scope, and how to offer content and paid access.
Ionela contributed vision, content, brand and user support. Ignacio led technical definition, architecture, development, integrations and evaluation. Product definition and review were collaborative.
A view of what was built
Portal, content and paid access connect with the assistant and its controls. Alongside the product, procedures and tools let the owner and team prepare content, review results and continue operations.
Explore this part of the product ↗
02 / Engineering decisions
Business vision + joint product definition
Access, reports and experience
Context, continuity and boundaries
Connected services
Consent, retention and deletion
Study review, articles and images
Newsletters, social channels and MCP
Team approval; events subject to consent
Response, memory and model tests
Assisted legal review and synchronisation
↺ Review feeds back into instructions, filters, skills and tests.
The user assistant and working agents have different roles. MCP connections belong to the working environment. Command-based legal review requires team supervision.
The user assistant has no unrestricted tool access. Code controls context and output; instructions and filters bound responses. Product development and operations agents serve a different role.
Define what information may be stored, with what consent and how it is deleted. Minimisation, retention and cascading deletion are accompanied by extraction and fact-continuity tests.
Review responses, sensitive cases and memory quality when models change. Findings guide corrections to instructions, filters and tests.
Transparency, data protection and review of health claims for the intended use, GDPR, the AI Act and the German context. Decisions, risks and controls are documented.
Operations, automation and controls
Consent, minimisation, retention criteria and cascading deletion. Tests examine memory extraction and continuity of facts.
Information about AI interaction and boundaries, instructions, filters and contextual review of public claims.
Scope, owners, risks, controls, providers and changes connected with evidence. Documentation references ISO/IEC 42001; it is distinct from a passed audit or accredited certification.
Reusable procedures for content, analysis and technical work; service integrations and a custom Meta Ads MCP server. The team has tools prepared through its iterations and corrections. Active connections depend on the working environment.
Procedures for newsletters, blogs and social channels with defined editorial voice and image styles. A scientific-paper auditing skill was also developed; it is in the disabled-skills archive, rather than running continuously. The product owner and team direct and review publication.
Usage, campaign and conversion events with consent controls and documented client/server flows. Measurement informs decisions and content revisions within the defined event scope.
Within the commit workflow, the team invokes a command to review content against German rules and propose or apply corrections. This is supervised assisted review, not an automatic Git hook.
Purchase, access and communication workflows, scheduled tasks, and commands coordinating changes, documentation and working environments.
03 / How it is evaluated
Tests cover assistant behaviour and boundaries, memory quality and model migrations. Findings feed corrections that can be checked again. Management documentation connects scope, owners, risks and evidence; implementation continues to evolve and is not presented as accredited certification.
A live product with a portal, content, reports, an assistant with memory and payments. A working environment with agents, tools and reusable instructions supports operations and evolution. The product owner directs content and reviews what is published.
Live product · case documented as of October 2026
What we can bring to your project