Verónica: AI-Powered Tutoring System

RAG Driven Vector Database Course-Embedded Tutor 8-Language Support

Verónica is an expert AI-powered tutoring system created by Enabling Learning. Embedded in every course and lesson, she is trained on expert-curated knowledge bases covering bilingual education, ESL, special education, and teacher certification content. Students can ask questions, reframe information, and receive scaffolded guidance within their zone of proximal development.

What Verónica Does

Four roles in one tutor. She teaches inside the lesson, evaluates spoken and written work against official rubrics, coaches the weakest skills first, and turns each learner's data into a plan.

Study Buddy

Embedded in every lesson, Verónica reads the actual page content, not a summary, so she can explain a concept in plainer terms, give a classroom example, generate practice questions, or pull out the key takeaways. Her answers are specific to the lesson the student is on.

AI Evaluator

She scores spoken and written responses against the official exam rubrics, task fulfillment, organization, and language use, and returns feedback that quotes the student's own words, names specific strengths and fixes, and shows a model response. Available in Spanish and English.

Smart Tutor

Verónica tracks a student's scores across every course they are enrolled in, coaches the weakest areas first, and recommends exactly what to study next, so practice time goes where it moves the exam score most.

Data-Driven Insights

She turns performance into a plan: a color-coded mastery heatmap across domains and competencies, a priority study plan that ranks weak areas by impact (domain weight × weakness), cross-course intelligence that flags concepts shared by several exams, and trend tracking that shows whether each skill is improving, flat, or slipping.

Conversation Practice

Students rehearse education topics out loud in the target language, and Verónica gives gentle corrections on grammar, accent marks, and academic register. Built for the Spanish BTLPT and LOTE, and for academic English.

Multilingual, by Design

Verónica works in eight languages, English and Spanish first, plus Mandarin, Vietnamese, Tagalog, Hindi, Urdu, and Brazilian Portuguese, in written text and natural-voice audio, so a student can learn in their first language. The set was chosen as the most common home languages of the students these teachers serve.

Model Lesson Plans

For the BTLPT, Verónica builds a model lesson plan from a real TEKS objective, showing the five required components: grade level, vocabulary, materials, procedures, and assessment.

Verónica is part of every Enabling Learning course, deployed across 16+ university and school-district programs, spanning the Spanish BTLPT (Bilingual Target Language Proficiency Test), academic English (TOEFL), LOTE, and 16 TExES certification exams.

Technical Architecture

How the RAG pipeline, embeddings, guardrails, and Vygotskian design work together under the hood.

The Problem with Generic AI

Why ChatGPT alone is not enough for education

Hallucination

General AI confidently invents facts about your specific course, policies, or exam requirements.

No Domain Expertise

Knows a little about everything, but nothing deep about bilingual education or your institution's content.

No Guardrails

Can go off-topic, give inappropriate responses, or be manipulated into ignoring its instructions.

No Pedagogical Design

Answers questions but does not scaffold learning, assess performance, or adapt to the learner's level.

The Solution: RAG (Retrieval-Augmented Generation)

Give the AI access to your own curated knowledge base. It retrieves relevant, verified information before responding. The facts come from your data, not its imagination.

What makes RAG truly powerful is embeddings, a technique that converts text into numerical representations of meaning, so the AI finds relevant content even when the question uses completely different words.

What Are Embeddings?

Turning text into numbers so computers can find meaning

Human Text

"How do I scaffold vocabulary for emergent bilinguals?"

→

Embedding Model

Converts meaning into a vector

→

Vector (Numbers)

[0.023, -0.851, 0.442, ...]

1,536 dimensions

Why does this matter? Embeddings capture meaning, not just keywords. "How do I scaffold vocabulary?" and "Strategies for academic word development in bilingual classrooms" have similar vectors even though they share almost no words. The AI finds knowledge by meaning similarity, not keyword matching.

Question Vector

[0.023, -0.851, ...]

≈

Knowledge Vector

[0.019, -0.847, ...]

→

Match!

Cosine similarity: 0.94

The RAG Pipeline

What happens when a student asks a question

Step 1

Student Asks

"How do I teach science vocabulary to ELLs?"

Step 2

Safety Screening

Regex + Moderation API filter

Step 3

Embed Question

Convert to vector (1,536 dims)

↓

Step 4

Vector Search (Filtered)

Search by meaning + filter by course context

Step 4b

Student Performance

With consent, retrieve scores & progress

↓

Step 5

LLM + System Prompt

Retrieved knowledge + student context + Vygotskian instructions

↓

Step 6

Personalized, Grounded Response

Based on verified knowledge, adapted to the learner's actual performance, within their ZPD

Four Layers of Safety

How we prevent hallucination, abuse, and off-topic responses

Regex Pattern Filtering

Catches abuse patterns and prompt injection attempts before the request reaches the AI model.

Moderation API

Screens every message for harmful, violent, or policy-violating content. A second independent filter.

System Prompt + Custom Code Restrictions

System prompt enforces behavioral rules (stay on topic, never invent information). Custom code restricts what topics the conversation can cover, limiting the AI to the specific educational domain.

Context-Aware Retrieval Filtering

Vector search is filtered by where the user is on the platform. A homepage visitor cannot access course content. A student in a lesson only gets relevant material.

The Knowledge Base

Expert-curated, not scraped. Verified, not generated.

850+ curated entries

15 content types

Products & Services Books & Research SPED Frameworks Glossaries Competencies ELPS Standards Scenarios FAQ

Each Course Gets Its Own Knowledge Filter

Homepage→ Company info, products, pricing, FAQ

BTLPT (Spanish)→ Grammar, academic vocabulary, proficiency strategies

Bilingual Ed.→ ELPS, bilingual pedagogy, competencies

ESL Supplemental→ ESL methods + shared bilingual theory

SPED EC-12→ TExES 186, 120 SPED terms, IEP

TOEFL / Acad. Eng.→ Academic English, exam strategies

IP Protection: Expert-created content is only retrievable inside the course environment. Casual visitors cannot access it.

The Vygotskian Design

How RAG enables AI tutors to function as mentors in the ZPD

Beyond Current Reach

Zone of Proximal Development

What learner
can do alone

RAG Enables Vygotskian AI

Expert Knowledge (MKO)

is implemented as

Curated Vector Database

Scaffolding

is implemented as

System Prompt Design (ADG model)

ZPD Assessment

is implemented as

Dual-Axis Evaluation (pedagogy + language)

Gradual Release

is implemented as

Adaptive Feedback (not answers, guidance)

Individualized Instruction

is implemented as

Performance-Aware Context (with student consent)

Swipe or use arrows to explore

Human Expertise Powered by AI

Verónica doesn't just process algorithms; she is fueled by the foundational knowledge and practical classroom wisdom of experienced bilingual educators and instructional designers. This human-in-the-loop approach ensures every response is pedagogically sound.

Interested in This Kind of System?

I design and build AI-powered learning systems grounded in research, pedagogy, and real classroom experience. If your institution or organization is exploring how AI can support teaching, learning, or professional development, I would love to talk.

Get in Touch Learn More About Verónica