Filip Pokorny

Led development of a multi-agent chatbot for financial planning tasks (budget forecasting, invoice management, cash flow analysis) fully integrated with Sage Intacct. Fine-tuned GPT-4 and BERT on financial documents to improve domain relevance. Implemented Retrieval-Augmented Generation (RAG) with a knowledge base using FAISS and Pinecone; integrated LangChain and GPT-4 for up-to-date insights, reducing response time and increasing accuracy. Added ReAct-style reasoning for multi-step financial planning tasks and multi-modal GraphRAG for processing reports and charts. Maintained MCP-based state management and deployed continuous fine-tuning with human feedback and RLHF. Operated production-grade MLOps with AWS SageMaker, EKS, and Docker, achieving high uptime and scalable usage. Evaluated model performance with LangSmith and OpenAI Evals and conducted A/B tests. Also developed an AI-powered fraud detection system for Sage 50 using anomaly detection in TensorFlow and AWS Lambda.

Led development of a multi-agent chatbot integrated with Sage Intacct for financial planning tasks such as budget forecasting, invoice management, and cash-flow analysis. Pre-trained and fine-tuned GPT-4 and BERT on financial documents and client queries, improving response accuracy by 25%. Implemented Retrieval-Augmented Generation (RAG) with a Knowledge Base using FAISS and Pinecone, and integrated LangChain, Langgraph, and GPT-4 to deliver up-to-date financial insights while reducing response time by 30%. Added ReAct for multi-step cash-flow reasoning and used CoT prompting to improve structure and user satisfaction by 20%. Extended to multi-modal retrieval (GraphRAG) with CLIP for visual data in reports, achieving 35% higher decision-making accuracy. Built continuous fine-tuning and RLHF pipelines with PPO/GRPO, automated retraining with AWS SageMaker and CircleCI, reducing error rates by 20%. Deployed at scale with AWS SageMaker, AWS EKS, and Docker; monitored with Prometheus, Gra

Led development of a multi-agent chatbot for financial planning tasks integrated with Sage Intacct; pre-trained and fine-tuned LLMs (GPT-4, BERT) on financial documents; implemented Retrieval-Augmented Generation (RAG) with FAISS and Pinecone; integrated ReAct for multi-step reasoning; added multi-modal GraphRAG with CLIP; built continuous fine-tuning and RLHF pipelines; deployed with AWS SageMaker, AWS EKS, and Docker; monitored with Prometheus, Grafana, and CloudWatch; achieved 99.9% uptime and improved user retention; evaluated with LangSmith and OpenAI Evals. Also developed an AI-powered fraud detection system for Sage 50 using anomaly detection and real-time inference.