GermanGuess

Background

As a programme lead in Germany I need B1 German for daily work. Existing apps do not cover my learning path. I built GermanGuess to understand where LLMs help with structured validation — and where they do not.

Problem

Professionals learning German need consistent, trustworthy practice — the same quality bar applies when scaling exercises without heavy manual review.

Goals

• Grammar exercises with immediate, reliable feedback
• Hybrid validation: rules + LLM variation
• Measurable learning curve across sessions
• Understand where LLMs fail in learning products

Architecture

Technology stack

React · TypeScript · Python · LLM APIs · PostgreSQL

Engineering challenges

• Reliably generating grammatically correct exercises
• Catching hallucinations and incorrect corrections
• Modeling learning progress meaningfully in PostgreSQL
• Keeping feedback loops faster than traditional learning apps

Trade-offs

• LLM variation vs. reliable rule checking — hybrid validation required
• Feature scope deliberately limited in favour of exercise quality
• Single-user focus — no multi-tenant architecture

Lessons learned

• AI works well for variation, less for reliable rule checking without additional logic
• Feedback loops must be faster than in traditional learning apps
• Personal projects make LLM limitations most visible

Current status

Live for users. Gathering feedback. Building additional learning modes.

Future improvements

• Extended grammar modules (subjunctive, passive voice)
• Better hallucination detection in LLM feedback
• Visualise learning progress and target weak areas for repetition

Estimated milestones

• Q2 2026 — Subjunctive module with rule validation
• Q3 2026 — Hallucination guardrails for correction feedback
• Q4 2026 — Weak-area dashboard