EdTech AI

Socratic Learning Machine

Intelligent Tutoring System for Critical Thinking

A Socratic AI that does not just answer. It thinks alongside the student. Adaptive question escalation driven by Bloom's Taxonomy and Webb's DoK, built on a fully custom async pipeline.

The Socratic flow loop

Student Responds

input + context window

Safety & Spell check

LlamaGuard · async · non blocking

Cognitive Assessment

Bloom's stage × DoK level

Clarifying

Remember · DoK 1

Probing Assumptions

Analyze · DoK 2 to 3

Synthesis / Evaluate

Create · DoK 4

Answer + Escalated Question

grade adapted · context rich · ↻ loops back

KhanMigo vs Socratic Learning Machine

KhanMigo Generic nudge engine

✕

One-size-fits-all follow-up

"Can you explain more?" regardless of cognitive stage
✕

No cognitive depth model

Does not map student to Bloom's or DoK level
✕

Static question repertoire

No escalation, same prompt pattern across all turns
✕

Flat retrieval

No curriculum graph, chunks retrieved by similarity only

This system Cognitive-aware Socratic engine

✓

6 targeted question types

Each mapped to a specific Bloom's stage and DoK depth
✓

Real-time cognitive assessment

Bloom's × DoK 2D map updated every turn
✓

Auto-escalating challenge

Clarifying → Probing → Synthesis as student improves
✓

Graph RAG curriculum awareness

Multi-hop retrieval across prerequisites and concepts

Socratic Learning Machine: FARM Stack Demo

FastAPI + React + MongoDB

Demo

Framework Custom · LangChain

Backend FastAPI · Python

Frontend React.js

RAG Graph RAG · Ontology

LLM Llama 3.3 70B · Qwen 3 32B

System architecture · click any node

Interface + Safety layer

Student interface

React · Pages, Components, Hooks

→

Safety & input validation

LlamaGuard · Spell check · Content filter

→

Conversation memory

MongoDB · Session context · Learning history

→

RAG retrieval to Graph RAG (v2)

PineconeDB · Jina AI 1024D · Ontology

→

Socratic flow engine (async)

Cognitive level assessment

Bloom's Taxonomy + Webb's DoK depth detection

→

Socratic question type selector

6 types · auto escalates with student progress

→

Response and follow up generator

Grade level adapted · context aware · async

→

LLM layer (agnostic)

Llama 3.3 70B · Qwen 3 32B · swappable

→

Graph RAG upgrade · v2 vision

Current: flat vector RAG

"Photosynthesis is the process by which..."

"Chlorophyll absorbs light energy at..."

"ATP synthesis occurs in the mitochondria..."

"Light reactions vs dark reactions..."

Chunks retrieved by similarity only. No concept relationships, no prerequisite awareness, no cross chapter linking.

V2: knowledge ontology graph

Textbook

Chapter 3

Chapter 7

Photosynthesis

Cell respiration

prerequisite →

Chlorophyll

ATP synthesis

explains →

Edges encode: prerequisite · explains · contradicts · extends.
Multi-hop retrieval across books, chapters, and topics.

Tech stack

Backend

FastAPI · Python

Frontend

React · JS

Database

MongoDB

Vector DB

PineconeDB

Embeddings

Jina-AI 1024D

Safety

LlamaGuard

LLM (primary)

Llama 3.3 70B

LLM (alt)

Qwen 3-32B