MONTY
Idle

Monty

Autonomous behavioral intelligence for early childhood classrooms

2,400+wiki pages generated autonomously
From ~150 ingested observation notes. Each note triggers behavioral assessment, node extraction, incident documentation, profile updates, and research retrieval. An equivalent manual effort would take ~150 hours of skilled documentation work: reviewing child history, categorizing behavioral patterns, writing structured records, and searching academic literature.
~60 min of human work per note × ~150 notes = ~150 hrs replaced

What this is

Monty is an AI agent that watches a Montessori classroom through teacher observation notes and builds a living, growing knowledge base of what it learns. Every note that arrives triggers a full reassessment: the agent reads the child's history, extracts behavioral patterns, updates a shared knowledge graph, and decides whether it needs to go find research papers to fill gaps in its understanding.

The knowledge graph is the core. Behavioral patterns like "frustration after error correction" or "peer-recruiting during sensory play" emerge as nodes that grow stronger with each supporting observation. Edges form between them as the agent discovers relationships. When enough children exhibit a pattern, or when something surprising appears, a curiosity gate fires and the agent autonomously searches OpenAlex for peer-reviewed literature to ground its observations in science.

Everything the agent knows lives in markdown files on disk. The database is a derived index that can be rebuilt at any time from those files. This means the agent's knowledge is human-readable, version-controlled, and auditable down to a single observation.

What you're looking at

This dashboard makes the agent's work visible in real time. You can watch behavioral patterns emerge as the force-directed knowledge graph grows on the Live page, read the actual markdown the agent writes on the Wiki page, and trace every processing step on the Console.

Live
Force-directed behavioral graph, student timelines, and the agent pipeline stage rail. The operational view.
Wiki
File-tree browser of the markdown knowledge base, inspired by Andrej Karpathy's LLM-wiki concept. Every behavioral node, student incident, and research paper the agent has written, browsable with backlinks and graph connections.
Console
Cycle state, throughput metrics, trace logs, and the curiosity event stream. The diagnostic view.
God Mode
Full operator control. Steer persona sliders, inject scenarios, trigger story presets, adjust curiosity sensitivity, and purge-restart the entire system.

The simulated classroom

Five synthetic children attend this classroom, each with a distinct temperament and stress response profile. A persona engine generates observation notes through an LLM, producing realistic Montessori-style documentation that varies with each child's current regulatory state.

Operators can adjust each child's regulatory slider from calm to dysregulated, change their stress response type, inject specific scenarios (a calm morning, an escalation, an emergency), and force peer interactions between children. The agent processes whatever comes through the pipeline and responds accordingly.

Why this matters

Early childhood teachers spend an estimated 8-12 hours per week on documentation and pattern recognition across their classroom. Monty eliminates the synthesis layer: reviewing notes, updating developmental records, and connecting observations to research. That returns 4-6 hours per teacher per week to direct instruction.

Months to minutes
Cross-child pattern recognition that takes 2-3 months of experienced observation surfaces in hours as the knowledge graph grows.
26% annual turnover
Average early childhood teacher turnover rate. Each departure loses months of child context. A persistent wiki eliminates institutional knowledge loss entirely.
Center for the Study of Child Care Employment, UC Berkeley, 2020
$5K-$15K per child
Published ROI of early intervention when developmental concerns are identified 3-6 months sooner. Real-time severity tracking and trend alerts accelerate identification.
RAND Corporation, "Investing in Our Children," 2005; Heckman, J., "Skill Formation and the Economics of Investing in Disadvantaged Children," Science, 2006
Zero to automatic
Research retrieval that teachers never have time to do. The curiosity gate autonomously fetches peer-reviewed papers from OpenAlex filtered to early childhood education topics.

This scales to organizations. A single classroom has 25 children. A school district has thousands. The same architecture maintains a unified, queryable knowledge base across all of them. Any authorized educator can ask "which children showed regression this month?" and get an answer grounded in aggregated observations, linked behavioral patterns, and cited literature.

The compound effect is the hardest to quantify and the most valuable. After a year of operation, the knowledge graph contains thousands of observations, hundreds of behavioral patterns, and dozens of research links. That institutional knowledge asset has no equivalent in a paper-based system.

How it works

Persona Engine
generates observation via LLM
1
Streamer
inserts into ingested_observations
2
Behavioral Assessment
LLM extracts severity + facets
3
Node Extraction
triggers, behaviors, brain states, responses
4
Wiki Writer
incident page + behavioral KG update
5
Student Profile
reassess full history, update summary
6
Curiosity Gate
score novelty, recurrence, surprise
7
Research Fetch
OpenAlex papers if score ≥ 0.70
8
Alert Generation
recommended actions for educators
9
during idle cycles
Idle Research
discover edges between disconnected nodes
10
Pushing the frontiers for Harnesses
Python + FastAPI + Next.js + SQLite + OpenAI + OpenAlex