Michael Accetto

Built Agentic Graph-RAG systems using Neo4j, custom MCP gateway, n8n orchestration and SolidStart frontend for energy-sector clients, integrating account data and ISO APIs for customer facing solution. Designed multi-agent workflows using n8n, specialised context engineering with custom ontology for lead generation, comprising of human-in-the-loop validation and source attribution. Deployed monitoring infrastructure tracking model behavior with Kubeflow in production with anomaly detection and automatic retraining triggers when market conditions shift.

Fine-tuned Gemma 2B and Phi-2 models on domain-specific data in GCP, achieving 8% accuracy improvement over base model with W&B tracking and automated evaluation pipelines. Zed IDE coding agent running on-device using quantized models for cost-effective inference. Designed and deployed AI-based CRM to automate comms with sales, outreach and service providers.

Fine-tuned Llama-2 7B and Falcon-7B on 250K+ proprietary Dutch legal documents using distributed training on GCP VM with nvidia A100. Implemented RAG system using chromadb for cosine similarity search with LangChain and streamlit. Developed LabelStudio interface and coordinated labeling of data through Vertex AI HITL Platform.