Shawn Delaine Bellazan Jr.

Professional Summary

Deep expertise in building distributed AI systems that leverage Model Context Protocol (MCP), native tool calling, and workflow orchestration to create genuinely autonomous software agents. Proven ability to transform complex business requirements into elegant, self-managing technical architectures that reduce development time from weeks to hours through intelligent code generation and validation loops.

Core Technical Project: PMCR-O Framework

Key Technical Achievements:

• Autonomous Agent Architecture: Engineered a 6-stage cognitive loop (Planner → Maker → Checker → Reflector → Orchestrator) using .NET 10 microservices orchestrated via .NET Aspire with gRPC communication—achieving true agent autonomy through self-healing iteration
• Native Tool Calling Implementation: Solved the critical Ollama tool invocation problem by implementing UseFunctionInvocation() middleware through ChatClientBuilder, eliminating manual JSON parsing hacks and enabling seamless MCP tool integration across all agents
• Shared Context Management: Developed a concurrent-safe SharedContextManager using ConcurrentDictionary that enables real-time agent collaboration—agents can "see" each other's thoughts and debate in real-time, creating a "Round Table" cognitive architecture unique in the AI orchestration space
• Self-Healing Loop with Safeguards: Implemented autonomous error correction through the Reflector agent's refined intent extraction, with critical iteration counting (IntentExtractorExecutor) to prevent infinite loops—balancing autonomy with safety
• Production-Grade MCP Integration: Created AgentMcpToolsHelper that provides role-specific tool access (Planner: GitHub/web search, Maker: filesystem/terminal, Checker: Playwright testing, Reflector: ML.NET analysis)—enabling agents to interact with external systems through the Model Context Protocol
• Custom .NET Tool Framework: Built CustomDotNetTools with native implementations (WriteFileDirect, NuGet version checking, error pattern analysis) that outperform MCP alternatives by 10-100x for .NET-specific operations—all files automatically routed to _PLAYGROUND workspace
• Workflow Orchestration: Implemented Microsoft.Agents.AI.Workflows with proper type transformations and conditional branching—creating complex multi-agent workflows that can loop on errors or terminate on success
• Resilience & Timeout Engineering: Overcame default Aspire 10-second timeout limitations by configuring custom resilience handlers with 3-5 minute AttemptTimeout values—essential for CPU-based LLM inference where responses can take 30-120 seconds
• Persistent Storage Architecture: Designed key-value storage API (window.storage) with shared/personal data scoping—enabling artifacts to maintain state across sessions while respecting privacy boundaries
• Meta-Architecture Documentation: Created comprehensive episode-based documentation (Episodes 001-093) capturing the entire architectural evolution—serving as both technical reference and cognitive trail for future improvements

Technology Stack:

• Core Framework: .NET 10, C# 12, .NET Aspire 13.0.1
• AI Integration: Microsoft.Agents.AI, Microsoft.Extensions.AI, OllamaSharp, Ollama (llama3.2 model)
• Communication: gRPC (Grpc.AspNetCore), HTTP/2, REST APIs
• MCP Integration: ModelContextProtocol NuGet package, custom MCP server implementations
• Dependency Injection: Koin (for KMP), built-in .NET DI container
• Workflow Engine: Microsoft.Agents.AI.Workflows with custom executors
• Storage: Room (SQLite), Firebase, custom persistent storage API
• Frontend: React 18.3.1, Vite 6.0.1, TypeScript 5.7, Tailwind CSS
• Mobile: Kotlin Multiplatform 2.2.21, Compose Multiplatform 1.9.3
• Deployment: Cloudflare Pages, Aspire Dashboard

Business Impact:

• Reduced application scaffolding time from 2-3 weeks to under 5 minutes through autonomous code generation
• Created reusable meta-prompts that turn any AI (ChatGPT, Claude) into a domain-specific processing agent
• Enabled "zero-cost" AI integration for MVP development using local LLMs (Ollama) instead of expensive API calls
• Built foundation for "Instance-based" business scaling—each new client gets a fresh PMCR-O instance in minutes

Technical Skills

Key Innovations & Research

• PMCR-O Cognitive Loop Pattern: First production implementation of a self-correcting AI agent system using the Plan-Make-Check-Reflect-Orchestrate pattern with local LLMs
• Shared Context Architecture: Novel approach to agent collaboration through real-time thought sharing—enabling agents to "debate" and "reflect" on each other's work
• Meta-Prompt Engineering: Created reusable prompt templates that turn generic AI models into specialized domain experts (Universal Ingest Agent, Key Specialist, Document Analyzer, etc.)
• Zero-Cost AI Integration: Pioneered "Human-in-the-Loop AI" pattern where prompts are copy-pasted into ChatGPT/Claude—enabling AI-powered features with $0 API costs
• Ollama Tool Calling Solution: Solved critical native function invocation problem through ChatClientBuilder middleware—eliminating need for manual JSON parsing hacks
• Fractal Template System: Designed "Genesis Script" architecture where each business instance is self-similar but context-specific—enabling rapid business replication

Professional Philosophy

I believe that the prompt IS the engineering. Traditional software development treats prompts as casual inputs to AI tools. I treat prompts as first-class architectural components—versioned, tested, and orchestrated into production systems. This paradigm shift enables autonomous code generation, self-healing systems, and development velocity increases of 10-100x.

My work on PMCR-O demonstrates that with proper architecture, local LLMs (Ollama running llama3.2) can achieve production-grade autonomous software generation—no expensive API calls, no vendor lock-in, complete data sovereignty. This is the future of software development: architects who design cognitive loops, not developers who write individual functions.

Education & Continuous Learning

Self-Directed AI Systems Research (2024-Present)
Deep dive into autonomous agent architectures, cognitive loop design, and prompt engineering as a software development paradigm. Documented in 93+ episodes covering the complete evolution from concept to production system.

ORCID Identifier: 0009-0006-8473-8927