05/08/2026 updated


Senior Software Engineer / Solution Architect | Agentic AI | RAG, LLM | Java/Python/TypeScript/Go
About me
20 years on enterprise production systems and three years of AI engineering. Built SearchMap.AI: 8 services, 130+ MCP endpoints, dual orchestration. Staff at Kaleyra (~1M msg/day, 99.9% uptime) and BMC (Fortune 500 SaaS). Partita IVA, available for B2B contracts.
Languages
Project history
Built SearchMap.AI, a personal AI agent that runs my consulting business operations, from job market research to contract evaluation. Eight containerized services in Python, Java, and TypeScript, with dual orchestration under the hood.
- Two orchestration engines in parallel: OpenAI Agents SDK and LangGraph.
- 130+ MCP tool endpoints across 16 modules. The main server runs Python on FastMCP with Streamable HTTP and per-user session routing.
- There is also a Java MCP server on Spring AI, plus a TypeScript one with Zod schemas.
- Agents and reasoning models operate the whole system through any MCP-compatible client. Day-to-day, I drive it from a chat interface.
- Retrieval on Qdrant with Pydantic structured extraction.
- Classification builds a taxonomy for each project from facets like domain, work mode, and experience level. The orchestrator reads it as filter context.
- Research runs on Celery workers. Each task goes from search to classification, with progress streamed to the UI via Redis pub/sub.
- Production resilience: circuit breakers for LLM tool calls, and workflow checkpoints so jobs can resume after a failure.
Stack: Python, Java, TypeScript, FastAPI, Spring AI, LangGraph4j, Celery, PostgreSQL, Qdrant, Redis, Next.js 16, React 19.
Models: GPT-5.1 for the orchestrator agent, GPT-4.1 and GPT-4.1-nano for classification with escalation, GPT-4o-mini for relevance filtering. Local embeddings via sentence-transformers all-MiniLM-L6-v2.