Research & Initiatives

Moving from "Black Box" to "Glass Box." We build instruments to measure the physics of AI, and engineering frameworks to optimize it.

Select an Initiative

Explore our open-source datasets and foundational data structures.

Our focus is on raw utility and transparency in data engineering, shifting from opaque conceptual models to hard, measurable artifacts.

Accretive-Abstract-State-Tree (A-AST)

A content-addressable data structure written from scratch in C, designed to drastically optimize agentic ingestion.

  • Bare-Metal Efficiency: Bypasses the bloat of higher-level languages to provide a highly performant backbone for state tracking.
  • Structural Integrity: Built specifically to manage complex, accretive data flows with minimal memory overhead.

View Repository →

WikiLoL (Lists-of-lists) Epoch-20260503

A massive, 16.4 GB dataset compiling structured list data parsed directly from Wikipedia.

This dataset was built to provide clean, structured relational data for model training and evaluation. It gained rapid traction upon release, crossing 150 downloads within its first month of availability.


View on Hugging Face →

Stanza-TinyStories

A deeply structured linguistic dataset applying the Stanza NLP pipeline to the TinyStories corpus.

This artifact provides transparent, parse-tree level annotations for researchers studying how small models learn syntax, grammar, and narrative structures.


View on Hugging Face →

Stanza-Wikitext-2

The foundational Wikitext-2 dataset, fully parsed and annotated using the Stanza framework.

Released to support radical transparency in data engineering, giving researchers ready-to-use, highly annotated text to accelerate mechanistic interpretability studies.


View on Hugging Face →