The Manhattan Kernel – A Deterministic, CPU-Bound AI Runtime Engine

Raising $10,000active

A CPU-only AI runtime written in Rust that isolates LLMs to non-executable hypothesis generation, using a 3-tier memory to maximize local LLM avoidance.

Donate:Manifund

A CPU-only AI runtime written in Rust that isolates LLMs to non-executable hypothesis generation, using a 3-tier memory to maximize local LLM avoidance.

Donate:Manifund

Project Details

I am building Manhattan Kernel because current AI agents are broken. Leaving state management and execution to an LLM creates total non-determinism.

My project fixes this with a strict Rust state machine. The LLM is completely isolated. It cannot execute code or hold system state. It acts strictly as an untrusted operator that proposes graph mutations. Everything is verified inside a cryptographic Kernel Structure Graph (KSG) before anything happens.

The main concept is LLM Avoidance. The kernel has 3 layers of memory (Episodic, Semantic, Procedural). When the system solves a bug using the LLM, it saves the path as a schema. Next time, the Schema Search Engine solves it using local heuristic search. No LLM tokens needed for known domains.

### Proof of Work & Code Status

This is not an essay. I have a compiling Rust codebase with 25 green integration tests checking the full compile-repair loop. You can check my code and tests here:

https://github.com/lipcsiklaszlo5-lgtm/DrManhattan-Project

### Budget

- Local Hardware (Apple Silicon for local 3B-8B model testing & CPU benchmarks): $2,000

- API Costs (Testing cloud fallback models): $1,500

- Developer Runway (Living expenses to work full-time on this): $6,500

People

László Lipcsik

creator

Funding Details

Start Date: -
End Date: -
Expected Duration: -
Funding Raised to Date: -
Annual Budget: -
Monthly Burn Rate: -
Current Runway: -
Funding Goal: $10,000
Funding Stage: -
Fiscal Sponsor: -

Grants Received– no grants recorded

Add grant

Funding Asks

Manifund funding goal

Open

Minimum

$10,000

Ideal

$10,000

Discussion

LLászló Lipcsik13d

Hi @LeoGao, I just submitted The Manhattan Kernel to your microgrant program. It is a strict, CPU-bound Rust runtime architecture that treats the LLM purely as an isolated hypothesis generator, driving execution in graph-space to maximize local LLM avoidance. We already have 25 green integration tests in our open-source repo. I would love for you to check out our architecture!