✦Research & Development

Public briefs from the aime platform team.

These briefs describe what the aime platform achieves and the principles behind it. Full technical papers — including architectures, methods and results — are shared with partners and investors under NDA.

WP-011

Public Brief

Full paper under NDA

Replayable Planning: An Event-Sourced Architecture for the aime Planner Agent

The Planner Agent is responsible for the shape of a learner's instructional journey — what is taught, when, in what order, at what pace, and how assessment is interleaved. Today it operates as a plan-replacement system: every significant educational signal causes a full regeneration of the plan. As plans grow, that pattern becomes costly, opaque and difficult to extend to multiple collaborators. This brief outlines aime's shift to a replayable planning architecture, in which the Planner Agent emits typed educational decisions rather than complete documents, and the current plan is reconstructed by replaying those decisions over a periodic snapshot. The posture extends concepts already proven inside the aime Loom™ Workflow Engine — deterministic event histories, replay-based state reconstruction and first-class audit — and brings the same properties to instructional planning: incremental updates, full historical visibility, explainable plan evolution, and a foundation for human-in-the-loop governance, branching experiments and rollback. The event taxonomy, projection engine design, snapshot strategy, conflict-resolution rules and integration with the rest of the agent stack remain aime's proprietary IP and are shared with partners under NDA.

Request access →

WP-010

Public Brief

Full paper under NDA

Smarter on Smaller Models: Architectural Optimization for Educational AI

Frontier-scale models are routinely asked to perform work that is not language work at all — counting, sequencing, identifier management, schema construction and arithmetic. In a platform built around seven cooperating agents, that pattern compounds into significant unnecessary cost and a brittleness that small models cannot absorb. This brief outlines the architectural posture behind aime's platform-wide shift to smaller, education-tuned models: narrower agent responsibilities, deterministic software in place of language work, schemas shaped for compact models, and validation treated as a first-class layer. The result is a target 60–80% reduction in prompt complexity and LLM cost across the platform with classroom quality held constant. Specific prompt restructures, schema decompositions, repair strategies and small-model mitigations remain aime's proprietary IP and are shared with partners under NDA.

Request access →

WP-001

Public Brief

Full paper under NDA

Deterministic Structured Generation for Educational AI

Reliable structured output is a prerequisite for production AI in education, where downstream systems consume lessons, diagrams, assessments and agent calls as typed data rather than free text. This brief outlines why off-the-shelf JSON-mode approaches fail at classroom scale and what changes when structured generation is treated as a first-class platform guarantee rather than a post-hoc parsing problem.

Request access →

WP-002

Public Brief

Full paper under NDA

Leo: Efficient Reasoning Models for AI-Native Education

Leo is the intelligence layer of aime — purpose-built reasoning models tuned for the contours of teaching and learning, designed to operate in both connected and offline environments. This brief examines why education benefits from compact, education-specific reasoning rather than dependence on generic large models, and how Leo enables responsive tutoring, lesson generation and agentic workflows in low-resource settings.

Request access →

WP-003

Public Brief

Full paper under NDA

ThinkCache: Accelerating Multi-Step Educational Reasoning

Classroom AI must support dialogue, explanation, revision and multi-step problem solving without introducing latency that disrupts learning. ThinkCache is aime's performance layer for iterative reasoning, designed to reduce redundant computation and make sophisticated learning tasks viable on modest infrastructure. This brief discusses the role of reasoning-aware acceleration in lowering the infrastructure cost of high-quality educational AI.

Request access →

WP-004

Public Brief

Full paper under NDA

Kern: An Agent Execution Layer for Educational Workflows

Conventional agent frameworks carry overhead that is poorly suited to the realities of education — high token cost, brittle tool protocols and execution layers built for general business automation. Kern is aime's agent execution layer, purpose-built for the workflows education actually requires: lesson generation, misconception reasoning, content coordination and interactive response. This brief outlines the design principles behind a leaner approach to agents.

Request access →

WP-005

Public Brief

Full paper under NDA

Loom: Workflow Orchestration for AI-Native Learning

Teaching unfolds as a continuous workflow — planning, delivery, assessment, reflection and adaptation. Loom is aime's orchestration engine, responsible for coordinating long-running educational workflows, agent collaboration and system state across the learning experience. This brief explains why education requires orchestration as a first-class concern rather than as application-level glue.

Request access →

WP-006

Public Brief

Full paper under NDA

ThinkBook: Structured Knowledge Beyond Traditional RAG

Conventional retrieval-augmented generation produces shallow context, weak conceptual continuity and elevated hallucination risk — none of which are acceptable in a teaching system. ThinkBook is aime's structured knowledge layer, designed as the authoritative substrate for lessons, diagrams, tutoring interactions and content generation. This brief outlines why education needs structured knowledge representation rather than document retrieval.

Request access →

WP-007

Public Brief

Full paper under NDA

EduRule: Pedagogy-Aware Transformation of Knowledge into Learning

Correct information is not the same as good teaching. Effective learning systems must account for curiosity, misconception, sequence, cognitive load and narrative clarity. EduRule is aime's pedagogy layer, responsible for transforming structured knowledge into teachable content. This brief examines pedagogical abstraction as a distinct system concern — the bridge between knowing and teaching.

Request access →

WP-008

Public Brief

Full paper under NDA

Text-to-Diagram: Structured Visual Generation for Educational Clarity

Diffusion-based image generation is unsuitable for instructional diagrams, where clarity, layout, labels and information density matter more than aesthetics. aime's Text-to-Diagram system prioritises structured infographics over artistic synthesis, aligning visual output with pedagogical intent. This brief outlines the design rationale for treating educational visuals as a structured output problem rather than an image generation problem.

Request access →

WP-009

Public Brief

Full paper under NDA

aime Pack: Offline Distribution for AI-Powered Lessons

Access to high-quality AI learning is too often constrained by bandwidth, device capability and cloud dependency. aime Pack is the platform's portable delivery format for AI-powered educational content, enabling personalised, interactive learning experiences to operate across devices without continuous connectivity. This brief examines offline-first content packaging as a foundational requirement for education technology at global scale.

Request access →