Hey guys! I am an Infrastructure Engineer turned Systems Architect who has worked for most of the worlds largest banks and defence contractors. Today I am open sourcing a piece of Infrastructure I built to address alot of issues I am seeing with engineers trying to glue together multiple databases to suffice the needs of AI data consistency.
My concern and reason I built this system is I was seeing a lack of security and access concerns from the teams I was working with who were presenting AI applications.
The key with this system is the unified Tensor itself
```sql
-- Find users similar to Alice who are connected to Bob
FIND NODE user
WHERE role = 'engineer'
SIMILAR TO 'user:alice'
CONNECTED TO 'user:bob'
```
One runtime. One query language. One consistency model.
The other issue is security and caching. I've seen agents run away and API costs spiral. The Neumann cache does semantic similarity matching so you don't hit the API twice for "What is 2+2" and "what's two plus two". The vault uses AES-256-GCM encryption with graph-based access control. If an agent doesn't have a path to a secret node, it can't read it. Full audit logging on everything.
Auto-checkpoints before destructive operations with interactive confirmation. If something goes wrong, roll back to any previous state.
It's got distributed consensus with some weird geometric conflict resolution stuff (6-way classification instead of binary commit/abort), HNSW for vectors, and delta replication that gets 4-6x bandwidth reduction.
Named after von Neumann because he unified code and data. This tries to unify your data models.
Still early but it works. Feedback welcome, roast my architecture, tell me why this is a terrible idea.
Shadylukin•1h ago
My concern and reason I built this system is I was seeing a lack of security and access concerns from the teams I was working with who were presenting AI applications.
The key with this system is the unified Tensor itself
```sql
-- Find users similar to Alice who are connected to Bob
FIND NODE user
WHERE role = 'engineer'
SIMILAR TO 'user:alice'
CONNECTED TO 'user:bob'
```
One runtime. One query language. One consistency model.
*Benchmarks (M-series silicon):*
- 3.2M PUT, 5M GET ops/sec
- Vector similarity: 150us @ 10K vectors (13x vs brute force)
- Query parsing: 1.9M queries/sec
The other issue is security and caching. I've seen agents run away and API costs spiral. The Neumann cache does semantic similarity matching so you don't hit the API twice for "What is 2+2" and "what's two plus two". The vault uses AES-256-GCM encryption with graph-based access control. If an agent doesn't have a path to a secret node, it can't read it. Full audit logging on everything.
Auto-checkpoints before destructive operations with interactive confirmation. If something goes wrong, roll back to any previous state.
It's got distributed consensus with some weird geometric conflict resolution stuff (6-way classification instead of binary commit/abort), HNSW for vectors, and delta replication that gets 4-6x bandwidth reduction.
Named after von Neumann because he unified code and data. This tries to unify your data models.
Still early but it works. Feedback welcome, roast my architecture, tell me why this is a terrible idea.
*Links:*
- GitHub: https://github.com/Shadylukin/Neumann