Dachi Abashidze

At TBC Bank, I build end-to-end production RAG systems serving 20 business departments and 1,000+ active users. I own architecture across LangGraph, Databricks Vector Search, Lakebase, fine-tuned BGE-M3 embeddings, and Databricks Apps.

I've built multi-agent systems on AWS Bedrock AgentCore, including supervisor-collaborator architectures, custom MCP servers, session-scoped memory, guardrails, and Lambda tools. I also manage the full MLOps lifecycle, from CI/CD and MLflow model registry to node-level tracing, access controls, and production debugging.

For Georgian, a low-resource language, I built a production ASR system by fine-tuning NVIDIA FastConformer on 108 hours of domain-diverse audio. It achieved 12.32 WER in internal testing, compared with 49.29 for Azure, while reducing manual labeling effort through pseudo-labeling and automated data pipelines.

I've also delivered call summarization, contradiction detection, fraud detection, customer engagement scoring, and BI automation. My work combines practical ML, LLM systems, retrieval quality, cost optimization, and reliable deployment on AWS and Databricks.