DeutschMeister

Overview

A production-grade German learning web app: B2-level grammar reference, vocabulary management, collections and exercises — plus an AI tutor that explains my mistakes using my own grammar reference as its only source of truth. Built because I am the target user: my CV says “German: B1 in progress” and this is how it’s progressing.

The challenge

Language apps optimize for streaks, not for the grammar wall every intermediate learner hits. And when I added an LLM tutor, the challenge shifted: a tutor that invents grammar rules is worse than no tutor. The answer to a wrong exercise had to be grounded, cited, and measured — not vibes.

My role & the AI’s role

I defined the Clean Architecture layering — API / Application (CQRS with MediatR) / Domain / Infrastructure (EF Core 8) — and the curriculum content model. AI agents generated the bulk of the CRUD handlers, validators and React components; I reviewed everything that crossed a layer boundary, because that’s where AI code quietly cheats.

The AI tutor phase was almost entirely agent-built, but against contracts I set: a spec adapted to the repo’s reality, a retrieval quality gate that had to pass before any LLM call existed, and eval thresholds in CI that decide when the feature ships. The agent wrote the pipeline; I own the bar it has to clear.

Key decisions

• Clean Architecture + CQRS on a personal project: overkill on purpose — it keeps the domain testable and gave me a .NET 8 reference codebase I reuse at work.
• RAG over my own corpus, no framework — 18 grammar sections indexed in the same PostgreSQL (pgvector + full-text, reciprocal rank fusion), Anthropic API called directly. The tutor cites sections like [§konjunktiv] and links back to the reference; if retrieval doesn’t cover the question, it says so instead of inventing a rule.
• Evals as the ship gate, not an afterthought — golden datasets versioned in the repo, deterministic retrieval metrics (hit@4 ≥ 85%) plus LLM-as-judge for faithfulness and correctness, wired into GitHub Actions. New AI features stay gated until their baselines pass; the Grammatik-Training generator additionally puts every AI-made exercise behind a human accept/discard review.
• Honest degradation — no API key configured means the tutor tells you it’s off and the rest of the app works untouched. Budget circuit breaker and per-user daily limits keep the whole thing under ~5€/month.
• JWT auth with roles from day one; a learning app with accounts is still an app with accounts.

Results

Live on my own infrastructure, used daily by its most demanding user. The retrieval layer finds the right grammar section for 92.5% of golden queries before real embeddings even land (hit@4, 53-query golden set); the tutor and exercise generator roll out behind those eval gates. The B1 exam will be the real integration test.