AI Language Lab
For years I prepared bilingual and language teachers for high-stakes proficiency exams. The AI Language Lab is what that work became: a connected system for language learning where full online courses and real-time practice apps run on a single AI layer.
The technology only became practical recently. Large language models made it possible to embed a single tutor, Verónica, inside every course lesson and every app, grounded by retrieval-augmented generation so she follows the real exam standard instead of improvising. Real-time speech models let learners hold a spoken conversation with an AI character over a live audio connection. And because every answer is captured as data, the system can evaluate performance against official rubrics and decide what each learner should practice next.
Two full courses anchor the Lab, the Spanish BTLPT (the Bilingual Target Language Proficiency Test, required for bilingual teacher certification in Texas) and academic English (TOEFL), both with AI tutoring, AI evaluation, and digital credentials built in. Three practice apps extend them: BTLPT Speaking and Academic English Speaking for live two-way voice, and Ortografía Pro, which turns Spanish accentuation and spelling into a data-driven game. The courses run on WordPress and LearnDash; the apps are Python and React; one AI layer runs through all of it.
The System at a Glance
Full online courses anchor the Lab, the Spanish BTLPT and academic English (TOEFL), both with AI tutoring, AI evaluation, and digital credentials built in
Real-time practice apps extend the courses on a common Python and React stack: BTLPT Speaking, Academic English Speaking, and Ortografía Pro
University and school-district programs where Verónica, the RAG-grounded AI tutor, is deployed
Inside the Lab
Two full courses and three practice apps, all running on one AI layer. The courses teach and certify; the apps give learners somewhere to rehearse the hardest skills out loud and by hand.
The Courses
BTLPT Preparation Course
A complete online course for the Spanish BTLPT (Bilingual Target Language Proficiency Test), the state proficiency exam required for bilingual teacher certification in Texas, serving 16+ universities and school districts. AI-powered practice across all four domains (reading, listening, speaking, writing), conversation practice with Verónica, and a four-domain capstone that scores speaking and writing with gpt-4o against the official rubric. A blockchain-verifiable digital credential is issued automatically on demonstrated mastery.
Academic English Course
An online course for academic English proficiency exams (TOEFL and beyond), aimed at international students. Verónica tutors inside the lessons, grounded by retrieval over the course's own content and the official task families. AI evaluation returns rubric feedback on speaking and writing, and a Smart Study Dashboard turns each practice session into progress data the learner can act on.
The Practice Apps
BTLPT Speaking
Live spoken-Spanish practice for the BTLPT. The candidate always plays the teacher; the AI always plays the parent, principal, colleague, or student, across 15 role-play scenarios drawn from the five exam categories. Real-time conversation over a live audio connection, timed exam modes, and a post-session report with rubric feedback and a transcript.
Academic English Speaking
Real-time English speaking practice for international students preparing for academic English proficiency exams. Twenty conversation scenarios from academic life, advisors, professors, financial aid, peer mentors, plus twelve timed presentation prompts, with feedback grounded in the official task families and scoring criteria.
Ortografía Pro
Gamified Spanish accentuation and spelling. 1,598 expert-verified challenges across 15 game modes covering acentuación, tricky letter pairs, and writing conventions, with a complete rules guide, smart spaced repetition, Verónica's coaching after every round, and a 10-hour CPE digital credential earned only by demonstrating mastery.
The Technology Behind It
Four capabilities run across the whole system, the courses and the apps alike. Each one was a deliberate engineering decision, not a feature added for its own sake.
Verónica, Grounded by RAG
The same AI tutor runs through every course lesson and every practice app. She is grounded in expert-curated knowledge bases using retrieval-augmented generation: before she answers, the system retrieves the most relevant passages, the BTLPT framework, the academic-English task families and scoring criteria, or the RAE orthography rules, and gives them to the model as authoritative context. The knowledge base is scoped per program so content never leaks across exams. Verónica is deployed across 16+ university and school-district programs.
Real-Time Voice Conversation
The two speaking apps hold a live, two-way spoken conversation between the learner and an AI character. Audio streams directly between the browser and the OpenAI Realtime API over a WebRTC connection, and the API key is never exposed to the browser: the server mints a short-lived session token for each conversation. Latency is low enough that the exchange feels like a real interview or parent meeting, and the learner always holds the professional role while the AI plays the other person.
AI Evaluation of Speaking and Writing
In the courses, students record spoken responses and write essays that are scored against the official exam rubrics. A gpt-4o evaluator grades each one, carrying anchored scoring exemplars to keep it honest, and returns criterion-by-criterion feedback on a 100 to 300 scale. Spoken responses are first transcribed with a speech-to-text model chosen specifically because it does not hallucinate on accented, second-language speech, which would otherwise corrupt the score.
Learning Analytics and Data-Driven Support
Every interaction is captured as data and turned into support. A Smart Study Dashboard builds a live picture of each learner's mastery by sub-skill with heatmaps and trend lines; Ortografía Pro detects error patterns directly from the attempt log and synthesizes a weekly study plan; capstone gates open only once practice thresholds are met; and program leaders see cohort-level analytics for their students. This is how the system decides what each learner, including each BTLPT candidate, should practice next.
Technical Approach
These are engineered systems, not generic chatbots: custom code that orchestrates multiple AI models through their APIs, with encoded rubrics and safety guardrails. Highlights of the technical decisions behind the suite, on two complementary stacks.
Frontend
- ✓ Practice apps are mobile-first React and Vite single-page apps on a common utility-first design system
- ✓ Web Audio API for live recording indicators and waveform visualization during speaking sessions
- ✓ One-click single sign-on from the storefront into every course and app, with no second login
Backend
- ✓ Two stacks by design: the courses run on WordPress and LearnDash with custom PHP plugins for AI evaluation, capstones, and credentials
- ✓ The practice apps are standalone Python (Flask) services running under gunicorn on a Linux VPS, each isolated as its own service
- ✓ Server-side grading and metering, a Postgres data layer, JWT single sign-on, and storefront provisioning by purchase
AI & Data
- ✓ OpenAI Realtime API over WebRTC for live two-way voice, with server-minted ephemeral tokens so the API key never reaches the browser
- ✓ Retrieval-augmented generation over verified, per-program knowledge bases grounds every tutoring and feedback response
- ✓ gpt-4o evaluation against official rubrics with anchored exemplars, plus high-accuracy transcription chosen to avoid hallucination on second-language speech
- ✓ Per-skill mastery modeling, pattern detection from the attempt log, spaced repetition, and cohort dashboards
- ✓ Blockchain-verifiable digital credentials (POK) issued on demonstrated mastery, including a 10-hour CPE credential in Ortografía Pro
Verónica is the AI coach inside this system
Verónica is the same tutoring system that powers all of my AI-powered learning products. See how the architecture, knowledge base, and pedagogical design come together.
Learn More About VerónicaInterested in learning more about this project or discussing a similar system?
Get in Touch