Building the Future of
Intelligent Software

Unlike Transformers which scale at O(n²) attention cost, MinGRU uses a minimal gated recurrent unit with log-space computation. This means:

• Infinite context window without quadratic memory scaling
• Runs entirely on CPU — no GPU required
• 2M to 50M parameter models that fit comfortably in RAM
• Hidden state persists across calls — genuine long-term memory

# MinGRU update rule (simplified)
z_t  = sigmoid(W_z * x_t)          # update gate
h_t  = (1 - z_t) * h_{t-1} + z_t * (W_h * x_t)
# No reset gate = 33% fewer parameters vs standard GRU

Two independent AI agents debate before producing a final answer. The structured deliberation loop surfaces edge cases a single model would miss:

1. TinyCPUDevAI generates an initial response from local context
2. MentorAI reviews, challenges weaknesses, and proposes alternatives
3. CouncilOrchestrator runs N debate rounds until consensus is reached
4. A joint answer is synthesised with a confidence score

Both agents share a common ToolRegistry with 48+ tools — either agent can autonomously invoke any tool during deliberation.

Every AI-driven file write routes through a three-layer safety pipeline before touching disk:

If any step fails, the original file is automatically restored from backup. The SyntaxGuard runs AST validation before committing — a corrupt AI output can never reach your codebase.

A transparent security layer sits underneath every subsystem — callers never have to think about it:

• Prompt injection guard with BiDi control character blocking
• Two-pass AST alias resolution — catches import os as o; o.system(...) bypass attacks
• SHA-256 checkpoint verification on every model load
• Entropy-based corpus poison detection — adversarial training samples quarantined automatically
• Token-bucket sandbox rate limiter (10 runs / 60s)