Who.
Diren Kumaratilleke. Nineteen. International freshman at the University of North Carolina at Chapel Hill, majoring in Information Science. The four projects on this site — BTUT, Crystara (TCD-JEPA), NIV (Regenerationism), and the Participatory Data Estate (SGUNCCH) — are solo work, zero co-authors, shipped under open-source licenses in the same calendar year classes were in session. No lab, no advisor on the byline, no outside funding for the research. One desk.
The four paradigms, named.
- BTUT — a new substrate for complex multi-agent systems.
- TCD-JEPA (Crystara) — a new AI training paradigm past transformers.
- Regenerationism (NIV) — a new economic philosophy; NIV is its first operational instrument.
- Participatory Data Estate — a new framework for digital data governance; SGUNCCH is its first live deployment.
Working method.
I am monotropic by disposition and hyper-relational by method. In practice that means I hold one system at a time, at unreasonable depth, and the moves across systems are structural rather than topical — the reduction in BTUT is the same reduction in Crystara is the same reduction in NIV, just in different typefaces. The four primitives are not four hobbies. They are one shape in four domains.
Every claim on this site is traceable to an open repository, a measured benchmark, or a walk-forward-validated number. Where the claim is about a future direction it is labeled as such.
Principles, explicit.
- Reductions, not refinements. A primitive earns its place by making a problem cheaper by an order of magnitude, not by 10%.
- Transparency by construction. Every weight in NIV is named. Every approval in PDE is logged. Every module in Crystara is typed. Opacity is a design bug.
- Composability over completeness. Small primitives with well-typed interfaces beat monolithic platforms. The four primitives are deliberately narrow.
- Validation over claim. Walk-forward OOS on NIV. Constant 12-iteration convergence from 500 to 10,000 agents on BTUT. Persistent-homology diagnostics on Crystara. Public approval log on PDE. The numbers are cited from the repos, not invented for the deck.
Constraints.
One person. A student budget. A single workstation and a handful of borrowed GPUs. Teaching load elsewhere (I TA). These are not obstacles for this research program; they are the shape of the research program. If a direction is not solo-feasible, it is not one I work on. That filter is why the primitives came out horizontal.
What this site is not.
This is not a résumé. It is a serialized argument for a different shape of AI — built out of four primitives that compose into a single inference-time substrate. The résumé is downstream of the manifesto. The manifesto is what matters.
Where the work lives.
- BTUT — github.com/direncode/btut
- Crystara (tcd-jepa) — github.com/direncode/tcd-jepa
- NIV (regenerationism) — github.com/direncode/regenerationism
- PDE (sguncch / Project Bold) — github.com/direncode/sguncch