← Back to Work
Case Study · Provizraq Labs

Hyper-Personalized
AI Learning — Proviz

An end-to-end intelligent education platform: dynamic course generation, AI video lectures, adaptive learning paths, a personal AI tutor, job-aligned skill mapping, and a living knowledge base — all tailored to each learner in real time.

Dynamic Course Gen AI Video Lectures AI Tutor Adaptive Learning Job Alignment Notes & Slides AI RAG Knowledge Base Multimodal Models
Dynamic Courses AI Tutor Job Adaptive Video Lectures
PROVIZ
Learn smarter · Grow faster · Land further
Courses generated
1:1 AI Tutor per learner
+68% Retention vs. static
< 2min Course gen latency
Real-time Path adaptation

Why education needed to be reinvented

Most online learning platforms offer the same course to every learner, regardless of what they already know, how they like to learn, what job they are targeting, or how much time they have. The result is dropout rates above 90% on MOOCs and millions of learners stuck in content that is too slow, too fast, or simply irrelevant to their goals.

Provizraq Labs was founded on April 3, 2025, with one mission: make education truly personal, accessible, and AI-powered for every learner on earth. Proviz is the flagship product of that mission — a platform where no two learning experiences are ever the same.

"I already know Python. I don't need the 10-hour beginner course. I need the specific 40 minutes that gets me from where I am to where I want to be." — Early user interview

The breakthrough insight was that the content itself should be generated, not retrieved. Instead of searching a catalogue, learners describe their goal — and Proviz builds the exact course, lecture series, notes, and quiz set they need, on demand.


Seven pillars of the Proviz learning stack

Every capability in Proviz was designed to work together as a closed loop: content is generated, delivered, assessed, adapted, and continuously improved based on how each learner actually engages with it.

Dynamic Course Generator
Generates a full structured course — modules, lessons, prerequisites, and assessments — from a single learner prompt in under 2 minutes.
🎬
AI Video Lectures
Produces narrated video lectures with generated visuals, diagrams, and synchronized slide overlays tailored to the learner's current level.
📄
AI Notes & Slides
Auto-generates structured notes, visual slide decks, and summary cheatsheets for every lesson — editable and exportable by the learner.
🤖
Personal AI Tutor
A conversational tutor with full course context, memory of past sessions, Socratic question style, and live feedback on learner answers.
🧭
Adaptive Learning Paths
Continuously re-sequences content based on quiz performance, engagement signals, and time-on-task. Skips what you know. Deepens what you struggle with.
💼
Job-Aligned Learning
Maps every module to real job descriptions, skill gap analysis, and career ladders. Every lesson is tied to a concrete career outcome.
🧠
RAG Knowledge Base
A retrieval-augmented knowledge layer that pulls from books, papers, docs, and web sources to ground every generated lesson in verified facts.
📊
Learning Analytics
Real-time dashboards for learners and instructors: mastery heatmaps, engagement funnels, concept-level struggle detection, and progress forecasting.

A full course from a single sentence

The course generator is the most-used capability in Proviz. A learner types what they want to learn and what they already know — and within 90 seconds, they have a fully structured, paced, and personalized curriculum.

Under the hood, a multi-step LLM pipeline extracts the learning goal, queries the knowledge graph for prerequisite concepts, runs a skill-gap analysis against the learner profile, designs module structures, and generates lesson-level content with assessments — all in a single coherent workflow.

"Teach me transformer architectures for NLP. I know basic Python and linear algebra."
↓ AI generates in < 90s
Module 01
Attention Mechanisms
4 lessons~40 min
Module 02
Encoder Architecture
3 lessons~30 min
Module 03
Decoder & Cross-Attn
4 lessons~42 min
Module 04
Pre-training & Fine-tuning
5 lessons~55 min
Module 05
Modern LLMs in Practice
3 lessons~35 min
+ Assessment
Capstone Project
Auto-graded

Generator pipeline stages

  • Goal extraction and intent parsing
  • Prerequisite graph traversal from knowledge base
  • Skill gap diff against learner profile
  • Module structure design and sequencing
  • Lesson content + assessment generation
  • Difficulty calibration and pacing

What makes it personal

  • Skips concepts the learner already knows
  • Adjusts depth based on prior test scores
  • Selects examples from learner's domain
  • Adapts reading level and explanation style
  • Reorders prerequisites intelligently
  • Sets time estimates to learner's pace history

Narrated, illustrated, and level-matched lectures

Proviz generates video lectures for every lesson — not just text. A multimodal pipeline writes a narration script, creates synchronized visual diagrams and slides, and produces an audio-visual lecture with chapter markers, transcript, and concept highlights.

Critically, the same concept is explained differently for a beginner vs. an intermediate learner. The system detects the learner's level from their profile and prior assessment performance and adjusts the vocabulary, analogies, and depth of explanation accordingly.

AI Generated · 12 min
How Self-Attention Computes Relevance Scores
Module 01 · Lesson 2 · Intermediate level · Auto-captioned
Narration Diagrams Slides Transcript Quiz
01
Script
LLM writes narration calibrated to learner level and style
02
Visuals
Diagrams, charts, and animations generated per concept
03
Slides
Slide deck synced to narration timestamps
04
Compose
Audio-visual assembly with chapter markers and captions
05
Deliver
Streamed to learner with interactive transcript overlay

A tutor that knows your entire learning history

The AI tutor is a persistent conversational agent scoped to each learner. Unlike a generic chatbot, it has full memory of every lesson completed, every quiz score, every concept the learner struggled with, and every question they have asked before.

The tutor uses a Socratic method by default — asking guiding questions rather than giving direct answers — to build genuine understanding. It can switch to direct explanation mode when a learner is stuck, provide worked examples on request, and generate new practice problems calibrated to the learner's current mastery level.

🤖
Proviz Tutor
You got 2 out of 5 on the self-attention quiz. I noticed you're struggling with the query/key/value separation. Before I explain it — what do you think the "query" vector is trying to find?
👤
Learner
I think it's looking for similar words nearby?
🤖
Proviz Tutor
Good intuition — but not quite positional. The query doesn't care where words are in the sentence, only what they mean in context. Think of the query as a question: "What information do I need to understand myself?" Each key is an answer from another token. The dot product scores how relevant each answer is. Does that reframe it?

Tutor capabilities

  • Persistent memory across all sessions
  • Socratic questioning by default
  • Struggle detection from quiz patterns
  • On-demand worked examples and practice
  • Concept linking to prior lessons
  • Tone adaptation: encouraging, direct, concise

Under the hood

  • Long-term memory via vector + structured store
  • Course context injected at each session start
  • Mastery state read from assessment events
  • RAG retrieval for factual grounding
  • Confidence scoring on generated explanations
  • Feedback loop: learner ratings improve tutor

Every learner takes a different route

Static courses give everyone the same sequence. Proviz continuously re-sequences content based on real signals: quiz scores, time spent, replay rate on video segments, tutor conversation patterns, and explicit self-assessment from the learner.

If a learner aces a concept, the system compresses or skips it and accelerates to harder material. If they struggle, it adds scaffolding, inserts bridging lessons, and slows the pace. The path updates after every interaction — not just at the end of a module.

Expert
Basics Intro Advanced Research Capstone
Intermediate
Basics Core Hands-on Advanced Capstone
Beginner
Prereqs Basics Core Practice Hands-on

The adaptation engine is powered by a Bayesian knowledge tracing model per concept, updated continuously with new evidence. It tracks not just "did they pass" but how confident and how durable their understanding is, and schedules spaced repetition for concepts at risk of being forgotten.


Every lesson tied to a career outcome

Proviz goes beyond skill trees. Each learner sets a target role or career direction, and the platform maps every lesson, module, and assessment back to real job-market requirements scraped and indexed from live job postings, company engineering blogs, and industry skill frameworks.

A skill-gap analysis runs at enrolment and re-runs as the learner progresses. It surfaces exactly which skills are blocking a promotion or a career switch, and auto-adjusts the course curriculum to close those gaps first.

🤖ML Engineer
Python PyTorch Transformers MLOps ⚠ Distributed training ⚠
🧠AI Researcher
Linear Algebra Transformers RLHF ⚠ Paper writing ⚠ Evals ⚠
⚙️AI Engineer
Python RAG LLM APIs ⚠ Agent design ⚠
📊Data Scientist
Statistics SQL Pandas Causal ML ⚠

Green pills = skills covered in current path. Red pills = gaps Proviz auto-schedules next.


How all the pieces fit together

Proviz is built as a modular AI stack where every capability is a first-class service with its own evals, latency budget, and fallback path. The learner profile is the central data primitive — every service reads from and writes back to it.

Proviz Platform — Layer View
Learner
Goal input Quiz responses Engagement signals Tutor conversations
Profile Core
Knowledge state (BKT) Skill graph Session memory Career target
AI Services
Course Generator Video Lecture Engine AI Tutor Agent Notes & Slides Gen Assessment Engine
Adaptation
Path Re-sequencer Spaced Repetition Difficulty Calibrator Pacing Engine
Knowledge
RAG pipeline (BM25 + dense) Concept graph Job market index Source vault
Delivery
Web platform Mobile app Instructor dashboard Analytics API

Measuring what actually moves learners forward

Every AI capability in Proviz is evaluated continuously against learner outcomes — not just model benchmarks. The eval suite tracks content quality, tutor faithfulness, path efficiency, and real-world learning retention.

+68%
Retention vs. static courses
92%
Tutor factual accuracy
89%
Course relevance score
< 90s
Course gen latency
4.7/5
Learner satisfaction
78%
Path skip acceptance rate
Concept mastery after adaptive path vs. linear +34%
AI tutor answer faithfulness (RAG-grounded) 92%
Job skill gap closure after 4 weeks 61%
Video lecture clarity rating (learner-reported) 4.6/5

What I designed, built, and owned

  1. 01 Founded Provizraq Labs Pvt. Ltd. on April 3, 2025 and defined the product vision: hyper-personalized, AI-generated education accessible to every learner globally.
  2. 02 Designed and built the dynamic course generator pipeline — multi-step LLM workflow for goal extraction, prerequisite mapping, skill-gap analysis, module design, and content generation with sub-90s latency.
  3. 03 Architected the AI video lecture engine: narration script generation, visual diagram synthesis, slide sync pipeline, and adaptive explanation calibration by learner level.
  4. 04 Built the personal AI tutor agent with persistent long-term memory, Socratic question-style prompting, struggle detection from assessment events, and RAG-grounded factual responses.
  5. 05 Designed the adaptive learning path engine using Bayesian Knowledge Tracing per concept, real-time re-sequencing, spaced repetition scheduling, and difficulty calibration from engagement signals.
  6. 06 Built the job alignment system: live job-posting indexer, skill gap analyzer, career ladder mapper, and curriculum auto-adjustment to prioritize blocking skills.
  7. 07 Implemented the hybrid RAG pipeline (BM25 + dense vectors + cross-encoder rerank) for the knowledge base, powering the tutor and course generator with verified, cited source grounding.
  8. 08 Designed the AI notes and slides generator: structured note formatting, visual slide deck production, and exportable cheatsheet generation for every lesson.
  9. 09 Conducted learner and stakeholder research, shipped eval harnesses for content quality and tutor faithfulness, and instrumented learning analytics dashboards for continuous improvement.