Prakti – work orchestration for internship programmes

In many SMEs, interns are under- and over-challenged at the same time: in one team delegable work piles up, in the next someone sits without tasks – and nobody sees it. Classic task managers don't solve this, because the real problem is trust: Who may involve whom? How well must a task be described to be delegable? And how does a reliable activity record emerge at the end without turning the whole thing into a surveillance system?

Exactly this gap – between task management and people development, under GDPR and works-council conditions – is what Prakti set out to close. Seven product principles were fixed before the first line of code, among them: capacity is exclusively self-reported, behaviour stays private; no rankings, no scores; matching stays rule-based and transparent, and the final decision is always made by a human.

Prakti was built specification-driven: market, product and UX concept, screens and technical architecture were decided as documents before anything was built – every feature follows the chain specification → data → API → agent interface → UI → tests.

Implementation was carried out by an AI agent (Claude Code) in interplay with human decisions at every fork in the road: stack choice, scope cuts and product principles remained management calls, while execution ran largely autonomously – including tests, code review by a second AI reviewer, CI/CD and deployment.

Technically, Prakti is a modular monolith on Next.js 15 with a framework-free domain layer, PostgreSQL with row-level security per tenant (EU hosting), magic-link and Microsoft Entra login, and task intake directly from Microsoft Teams and Slack. Data protection is a product feature, not a footnote: no activity, presence or screen data; metrics only aggregated with a minimum group size.

After roughly one week, a publicly explorable demo stood with the complete core workflow: report capacity → post a task through the quality gate → cross-team approval → acceptance and execution → review → automatically generated, verified activity record as PDF/CSV.

Over 630 automated tests (unit and integration tests against real PostgreSQL instances) protect every change; deployment runs via GitHub Actions to Cloud Run with no standing server costs (scales to zero). A guided tour with real screenshots makes both roles tangible in under two minutes – as a programme manager on desktop and as an intern on a smartphone, each with one-click demo login.

At the same time, the pilot shows how we build software today: AI-assisted in execution, human in the decisions, specification-driven in quality.