Project
Gestalt — A Multi-Provider AI Workspace
A personal project in development: Python orchestration, MCP execution, React, and Rust/Tauri, with permissions separated from prompts and explicit recovery safeguards. Architected and directed with AI-assisted implementation.
Gestalt explores how an AI workspace can keep conversations, reusable prompt skills, and provider execution coordinated without treating a prompt as permission to act. It is a personal project in development, with AI-assisted implementation under my architecture and direction.
Architecture
The work spans six repositories: MCP provider execution, Python orchestration, React messaging, a Rust/Tauri desktop application, a Kotlin Android companion prototype, and offline training and evaluation. Persistent AI personas and group conversations share a workspace while presentation, orchestration, and provider execution remain separate concerns.
Execution and recovery
I directed the implementation of immutable execution plans, capability and billing checks, permission revocation, cancellation, and durable execution receipts. Uncertain outcomes remain unresolved instead of being silently retried. That distinction matters when repeating an operation could repeat a side effect.
Adversarial review and restart tests include Windows process-death scenarios with simulated provider execution, checking for duplicate dispatch. These are local tests, not evidence of production reliability with live providers.
Evaluation
Workflow selection considers policy, evidence, cost, and latency. Offline evaluation keeps dataset provenance and separate training, calibration, and held-out evaluation sets. The project documents what was tested and what remains unverified.