Scrunchee Systems | Mission Control

[ NEUMANN ] STATUS: ONLINE

Unified Tensor Runtime
"Relational + Graph + Vector. One runtime. No glue code."

A production-grade database runtime that unifies relational tables, property graphs, and vector embeddings in a single address space. No microservices. No network hops between engines. Query across all three models in one statement.

395K

Lines of Rust

Days to Build

33K+

Tests

Crates

3.2M

Writes/Sec

Reads/Sec

226

Commits

93%

Coverage

First commit: December 24, 2025

Three Engines, One Runtime:

Relational Engine: Columnar storage, SIMD-accelerated filtering, B-tree indexes, full transaction support
Graph Engine: CSR format, BFS/DFS traversal, shortest path, property graphs with typed edges
Vector Engine: HNSW index, 15+ distance metrics, sparse vectors, tensor compression

Infrastructure:

tensor_chain: Raft consensus, 2PC distributed transactions, semantic conflict detection
tensor_vault: AES-256-GCM encrypted secrets with graph-based access control
tensor_cache: Multi-layer LLM response cache (exact + semantic + embedding)

GitHub Repository Documentation

[ SCRUNCHEE ] STATUS: PAUSED

Code Intelligence Platform
The first application of the stack.

Code intelligence built on Neumann + tensor_chain. Transforms codebases into semantic knowledge graphs where every function, type, and relationship has geometric meaning. Multi-agent AI reaches consensus through coherence, not voting.

"Paused while we build the foundation right. The infrastructure comes first."

When Ready:

Code as knowledge graph - functions, types, dependencies as nodes and edges
Semantic search across repositories - find by meaning, not keywords
Multi-agent consensus - AI agents agree through geometric coherence
Traceable confidence - every answer shows which code it came from

[ TENSOR_CHAIN ] STATUS: FOUNDATION BUILT

Geometric Consensus
Proof-of-coherence, not proof-of-work.

"What if consensus was geometric, not economic? What if AI infrastructure belonged to everyone who maintains it?"

Most blockchains validate by re-running computation (expensive) or counting stake (plutocratic). Tensor_chain validates by checking if new information coheres with existing knowledge - mathematically, through embeddings.

Already Built (65K lines):

Geometric Consensus: Similarity-based fast path - blocks aligned with state skip full validation
6-Way Conflict Detection: Cosine (angular) + Jaccard (structural) classifies transactions as orthogonal, conflicting, identical, or opposite
Codebook Vocabulary: K-means centroids define valid states - the boundary of coherent knowledge
Delta Embeddings: Every transaction carries semantic meaning as a sparse vector, not just bytes
Semantic Partition Merge: 6-phase protocol reconciles split-brain using vector operations

The Vision:

A self-sustaining intelligence system where:

The AI maintains state by paying for storage/compute
Payment comes from the value of the world-model it's constructing
Novel contributions expand the codebook and get rewarded
Value recognized in its natural form - code, art, care, labor

The foundation exists. The economic layer is next.

[ OPERATOR ] ONLINE

Lukin Ackroyd
Auckland, New Zealand

I'm a systems architect. Self-taught, no CS degree. I've spent the last decade building enterprise infrastructure for organizations like NZ Defence Force, Z Energy, and the Australian Bureau of Meteorology.

Neumann started as an experiment in December 2025 and became a working runtime by January 2026. I built it using AI agents as collaborators, treating them like a small team rather than autocomplete. The architecture worked because I think in systems - I see how pieces connect before I see the pieces themselves.

Before tech, I led a 200-person gaming guild to top-tier competitive achievements. It taught me that complex systems are really coordination problems, and coordination problems are really communication problems.

I'm building Scrunchee Systems because I believe infrastructure should be understandable, not abstracted away. If you can't see what the machine is doing, you can't trust it.

Resume lukin@scrunchee.ai