Projects

Work exploring how cognition might work in machines—and tools to help build the systems that test those ideas.

LOGOS

Active

Non-linguistic cognitive architecture

LOGOS is an attempt to build autonomous agents that reason without depending on language. Where language models think in tokens and generate plausible-sounding sequences, LOGOS represents knowledge as graphs and reasons over structure.

The system uses Neo4j for causal knowledge graphs and Milvus for embedding-based perception. Plans are validated against explicit preconditions and effects rather than generated and hoped for. The agent maintains a model of its own capabilities and updates it as it learns.

Components

SophiaNon-linguistic cognitive core
HermesLanguage and embedding utilities
ApolloUI and command layer
TalosSensor/actuator abstraction

Supporting Work

Experiments, prototypes, and tools that feed into or support the main work

TinyMind

Prototype

Curiosity-driven knowledge graphs

A prototype exploring how meaningful graph structure can emerge from curiosity-driven exploration rather than predefined schema. Edges encode semantic relationships that the system discovers—applies_to, has_consequence, basis_for—and the interesting part is that these emerge in ways that wouldn't be hand-designed.

TinyMind served as an incubator for ideas that informed Sophia's learning architecture, particularly around how topology can carry information independent of node content.

JEPA→CLIP

Experiment

Bridging video understanding to semantic space

Experiments in translating V-JEPA video embeddings into CLIP's semantic space. V-JEPA captures temporal dynamics and motion patterns; CLIP provides language-grounded semantics. The goal is enabling non-linguistic systems to ground their perception in concepts that can be communicated.

Current results show the translation is non-trivial—the embedding spaces encode fundamentally different information—but partial alignment is achievable.

agent-swarm

Active

Workflow enforcement for AI coding agents

Tooling for orchestrating AI coding agents in Claude Code. Provides structured workflows with checkpoints, phase transitions, and enforcement mechanisms to keep agents on track rather than wandering through codebases.

genetic_hackathon

Experiment

Discovering math from behavioral signatures

A genetic programming experiment that discovers mathematical functions by evolving expressions from primitives. Functions are identified by what they compute—input/output pairs—rather than syntactic structure, enabling the system to recognize when different expressions are functionally equivalent.

The system discovered transformation rules autonomously, including logical equivalences like De Morgan's laws, by collecting behaviorally equivalent alternatives and learning simplification patterns.