Building the primitives the field assumes aren't possible yet.
Semantic compression, deterministic retrieval, and models you can program in the weights. Everything below has a working implementation and something public you can go look at.
Work
ShipItClean
Hostile AI agents assume your code is wrong and prove where. Adversarial security audits with exploit-level reasoning and severity-ranked findings, run in parallel over a whole repository and cross-validated against each other. Not a code helper — an independent validator.
BTI Engine
The deterministic retrieval layer underneath Atlas. Exact recall across unlimited knowledge bases — no embeddings, no approximation, and no degradation as the corpus grows.
jBlaze
Programs behaviours and knowledge directly into model weights rather than into a prompt. Block injection, implant knowledge, and keep the change after the context window is gone.
LoreTokens
An AI-native semantic compression format. Compress meaning, not bytes — and leave the compressed form directly addressable, with no decode-through-a-model step in the way.
SAIQL
A database and semantic query language purpose-built for AI systems. Governed deterministic retrieval, semantic search, agent memory and enterprise migration under a single stack.
Sharona
A relational foundational model that never forgets, paired with Carl — the open-weight family that runs on your own machine. When Carl reaches its limit, Sharona finishes the job.