Patrick.Spelsberg pspelsberg

🚀 Hi there, I'm Patrick (pspelsberg) 👋

⚡ About Me

I am a software engineer and AI systems architect specializing in Autonomous Agent Harnesses, Local-First Architectures, and Deterministic System Engineering. I build resilient bridges between next-generation frontier LLMs (Mistral, Claude, Gemini, GPT) and local open-weights intelligence (Ollama, LM Studio, DeepSeek, Qwen), backed by robust sandbox isolation, privacy vaults, and test-driven self-healing loops.

🧠 Current Flagships: Architecting HarnessForge (visual state-graph agent engineering) and MCP Enterprise Gateway (sandboxed FastMCP gateway with GDPR vault).
🤖 Custom Open-Weights Model: Creator of LottiCode (5B) on Hugging Face — specialized pure-execution coding engine fine-tuned for high-speed sub-agent task synthesis.
🛡️ Core Philosophy: Zero cloud lock-in, deterministic execution boundaries, Local Trust Mode, and prompt-as-a-variable workflows.
🧩 Architectural Focus: Vertical Slice Architecture (VSA), Modular Monoliths, OKF (Open Knowledge Format v0.2), and bounded vector retrieval.
🔭 Research Interests: Recursive Language Models (RLM), ARC AGI challenges, MCP tool governance, and real-time multimodal audio streaming.

🏛️ Architectural Principles

┌──────────────────────────────────────────────────────────────────────────────────┐
│                             ENGINEERING FOUNDATION                               │
├───────────────────┬───────────────────┬──────────────────┬───────────────────────┤
│  🏠 Local-First   │  🛡️ Sandboxed     │  🧩 Vertical     │  🔄 Test-Driven       │
│     & Agnostic    │     & Private     │     Slices (VSA) │     Self-Healing      │
│                   │                   │                  │                       │
│ Zero cloud lock-  │ GDPR/DSGVO RAM    │ High-cohesion    │ Autonomous TDD loops, │
│ in, standalone    │ vaults, unprivi-  │ domain slices,   │ bounded retries, and  │
│ python exports,   │ leged containers, │ zero spaghetti,  │ SWE-bench RAG         │
│ local-model first │ & PII masking     │ strict contracts │ context seeding       │
└───────────────────┴───────────────────┴──────────────────┴───────────────────────┘

🌟 Featured Public Ecosystem

⚡ HarnessForge

Visual state-graph & loop engineering for autonomous AI agents on localhost.

Local-First: 1-click standalone Python export (agent_runner.py) with zero lock-in.
AI Architect: Natural language prompt-to-graph synthesis via Mistral AI & Ollama.
LanceDB RAG: High-speed bounded vector retrieval & dynamic prompt bindings.
Security: chmod 0600 API vault, Local Trust Mode & CodeUltra Full Tier.

🌐 MCP Enterprise Gateway

Enterprise FastMCP gateway with GDPR privacy vault & sandboxed tools.

Privacy Slice: Presidio + Regex + RAM Vault for GDPR/DSGVO-compliant PII masking.
Knowledge Slice: LanceDB hybrid & vector search with OKF v0.2 resources.
Sandbox Slice: Unprivileged Docker execution (network: none, read_only).
Audit Studio: Zero-cookie, local single-user UI with real-time audit stats.

🧬 Model Fusion

Enterprise software manufacturing powered by dual-model fusion & CodeUltra security.

12-Step Workflow: Dual-model synthesis with adaptive "Grill-Me" security audits.
OKF v0.2 Graph: Open Knowledge Format representation with 3-zone architecture matrix.
Threat Intelligence: Perplexity Sonar live threat research & contract validation.
Memory & Concurrency: Anti-timeline LanceDB vector memory & GPU locking queue.

🍣 miniSashimi

Intelligent multi-agent proxy for coding agents with real-time glassmorphism dashboard.

7 LLM Providers: OpenAI, Anthropic, Google, Mistral, OpenRouter, DeepSeek & LM Studio.
13 Configurable Roles: Granular assignment of workflow phases to specialized models.
Boundary Anonymizer: Runtime PII masking & AES-128/Fernet key storage (PBKDF2 600k).
Live Dashboard: Dark Glassmorphism UI with real-time SSE streaming & cost limiting.

🎙️ VoiceTL

Real-time bidirectional live voice translation agent for meetings.

Live Translation: Dual-stream voice translation for Slack Huddles, Zoom & MS Teams.
Ultra-Low Latency: Powered by gemini-3.5-live-translate-preview over WebSockets.
PipeWire Routing: Automated Linux null-sink loopback management (pw-loopback).
Cost Protection: Integrated RMS Noise-Gate VAD with real-time terminal VU meter.

🌌 Mini-CLI

Autonomous multi-agent AI coding assistant coordinating 21 specialized sub-agents.

Multi-Agent Swarm: Central orchestrator coordinating 21 specialized agent roles.
SWE-bench RAG: LanceDB context database seeded with SWE-bench benchmarks.
Autonomous TDD: Automated test-driven generation, validation & self-healing loops.
Interactive REPL: Rich terminal UI with live token telemetry & multi-provider routing.

🤖 Open-Weights AI Models & Fine-Tuning

🐒 LottiCode (5B) — Pure Execution Code Model

"Zero architectural bloat. Maximum 16-bit execution precision. Built to code down exact specifications."

Specialized Role: High-speed code-monkey worker model engineered to execute bite-sized, deterministic implementation tasks received from an architecture lead agent.
Execution Stack: Pure generation across TypeScript/JavaScript, Express API routes, Python backends, and Tailwind CSS utility layouts.
Local Hardware Optimization: 100% VRAM offload compatible with LM Studio & Ollama (Vulkan/ROCm acceleration on Linux). Compact context window (8192) for maximum output precision.

🛠️ Tech Stack & Ecosystem

🤖 AI, Multi-Agent Swarms & Vector Memory

💻 Frontend, UI & Real-Time Visualization

⚙️ Systems, Security & DevOps

📊 GitHub Analytics & Activity

_{Forged with precision. Connect with me on GitHub.}

Navigation Menu