I design the layer between an LLM demo and a system people can trust.
from dataclasses import dataclass
@dataclass(frozen=True)
class AIEngineer:
name: str = "Georgios (Giorgos) Agrafiotis"
role: str = "Senior AI Engineer · AI Solutions Architect"
experience: str = "8+ years building production software"
mission: str = "Turn ambitious AI ideas into governed, observable systems"
def expertise(self) -> dict[str, tuple[str, ...]]:
return {
"agentic_systems": ("LangGraph", "MCP", "multi-agent workflows", "memory"),
"ai_governance": ("OAuth/OIDC", "RBAC", "tenant isolation", "tool access"),
"platforms": ("AWS Bedrock", "Azure AI Foundry", "Google Vertex AI"),
"backend": ("Python", "FastAPI", "asyncio", "Go", "PostgreSQL"),
"production": ("OpenTelemetry", "Langfuse", "Kubernetes", "Terraform", "CI/CD"),
}Agents reason. Software enforces. Traces tell the truth.
Public stories from private repositories. The outcomes are visible; confidential implementation details stay private.
Quicklizard · Private repositories
Designed and built the AI layer around a dynamic-pricing platform.
| 🔌 Custom Pricing MCP Internal capabilities become structured agent tools |
🛡️ MCP Gateway Governed access for users and machine clients |
🧠 Pricing Agent Stateful, streaming pricing intelligence |
🔭 Observability Model-to-tool-to-service operational visibility |
Gresham Tech · Private repository
Designed and implemented a durable exception-management system for financial-data operations. It coordinates intake, resolution workflows, external interactions, recovery paths, and human review while keeping workflow state explicit and auditable.
| 📥 Intake Structured exception admission |
🔁 Stateful Workflows Explicit, recoverable execution |
🌐 External Actions Controlled system interactions |
🧑⚖️ Human Review Auditable escalation paths |
XGS.AI · Private repository
Took an existing RAG proof of concept to production: redesigned the service architecture, built the FastAPI and LangGraph integration layer, moved model and database paths to async workflows, and added conversation memory, delivery automation, and observability.
| 🏗️ Architecture POC to production service |
⚡ Async Core Model and database workflows |
🧵 Memory Conversation continuity |
📡 Operations Delivery, tracing, and alerts |
Technical assessments, applied prototypes, and ideas worth testing in public.
|
Cross-Source Investigation Assistant A checkpointed LangChain agent correlates evidence across comms, transactions, and documents, answering questions with cited, traceable answers. Explore repository → |
LangGraph E-Commerce Agent Routes questions, generates and retries BigQuery SQL, synthesizes insights, and keeps conversational context. Explore repository → |
Hybrid AI + Deterministic Logic An LLM extracts vessel data; typed Python owns every auditable financial calculation. Explore repository → |
| APPLIED R&D PROTOTYPE | TECHNICAL ASSESSMENT | TECHNICAL ASSESSMENT |
Thessaloniki, Greece · Open to ambitious AI systems and difficult production problems