RUTHRAN RAGHAVAN
Case Study|Language Education

Épreuves — DELF B1 Exam Simulation Platform

A self-hosted DELF B1 exam simulator with local-LLM written evaluation and an AI oral examiner.

The Problem

Understanding the business challenge.

DELF B1 candidates have limited access to realistic, full-length practice exams with immediate feedback.

Written and oral production sections require subjective evaluation that's hard to scale without human graders.

Sending exam responses to third-party cloud AI raises data-privacy concerns for an education provider.

The Solution

Our AI-powered approach.

A self-hosted exam platform simulating all four DELF B1 sections, using a locally-run LLM for written-response evaluation and an AI examiner persona with natural French text-to-speech for the oral exam — keeping candidate data fully on-premise.

Full DELF B1 Simulation

All four exam sections — listening, reading, written production, and oral production.

Local LLM Evaluation

Written responses are graded by a self-hosted Ministral 3B model — no candidate data leaves the server.

AI Oral Examiner

An interactive spoken interview with an AI examiner persona, powered by MLX-optimized Kokoro TTS for natural French speech.

Candidate & Admin Dashboards

Separate portals for exam-takers and administrators.

Single-Container Deployment

Postgres, backend, LLM, and TTS are bundled into one Docker container for straightforward self-hosting.

Technical Architecture

Enterprise-grade technology stack.

Frontend
React 19 with Tailwind CSS and Vite
Backend
Node.js and Express with PostgreSQL
Vector Database
Not applicable — direct LLM evaluation, no retrieval system
Models
Self-hosted Ministral 3B (Q4_K_M, via llama.cpp) in production; Qwen2.5:3b via Ollama in development
Retrieval
N/A — written responses are evaluated directly by the local LLM against exam rubrics

Security & Compliance

Privacy-first implementation.

  • Fully self-hosted LLM evaluation — no candidate responses sent to third-party AI providers
  • Session-based authentication with bcrypt password hashing and rate limiting
  • Single-container deployment minimizes external attack surface

Implementation Workflow

Structured deployment process.

1

Exam Content

Digitized all four DELF B1 exam sections with grading rubrics.

2

Local LLM Integration

Wired written-response evaluation to a self-hosted Ministral 3B model.

3

AI Oral Examiner

Built an interactive spoken interview flow with Kokoro TTS for natural French audio.

4

Containerized Self-Hosting

Bundled the database, backend, LLM, and TTS engine into a single production container.

Business Impact

Measurable outcomes.

Gives DELF B1 candidates realistic full-exam practice with immediate written and oral feedback
Keeps all candidate exam data on-premise by avoiding third-party AI APIs entirely
Reduces the grading bottleneck for written and oral sections that traditionally require human examiners

Target Industries

Versatile application across sectors.

Education
Language Learning