Ethan Domokos

• Led design and development of an enterprise agent that lets CRM teams build audience segments over first-party data enriched with multiple ML model attributes, then push them straight to platform for marketing activation — entirely through natural language.
• Productionized a suite of GenAI applications, including an LLM-driven semantic-model agent for governed natural-language querying and visualization, and a persona-simulation agent that synthesized market research into interactive AI personas, cutting manual focus-group analysis time through on-demand qualitative simulations.

• Orchestrated a foundational data ecosystem and governance framework for a high-profile public-sector client. Led a team of three to deploy 200+ automated ETL/ELT pipelines processing 40B+ records, migrating legacy silos into a unified BigQuery environment for real-time analytics.
• Built the orchestrated pipeline behind a data product consumed nationally, replacing the previous production process and saving $1.4M annually in operating costs.
• Architected a Bayesian Small Area Estimator within that pipeline as a containerized serverless service, mitigating market concentration and data suppression. Engineered geographic proximity anchoring to rebalance category and origin market spend distributions, doubling unique actionable data series.

• Built an end-to-end MLOps solution to improve GA4 user experiences and paid media activation, using machine learning to generate user attributes for advanced audience segmentation.

• Migrated legacy BI ELT pipelines off Power BI’s Power Query into custom BigQuery orchestration across 10 active client dashboards, cutting refresh time from 12 hours to 10 minutes and lifting refresh uptime from 70% to 99%.

Neuromorphic long-term memory for AI agents: brain-inspired persistent context shared across Copilot, Claude, and Cursor.

• Scores 98.2% Recall@5 on LongMemEval, beating GPT-4o-based systems using only local 22M-parameter models, with flat ~130ms retrieval latency from 100 to 1,000 memories via HNSW indexing.
• Zero-LLM architecture: recall, classification, and semantic operations all run on local models, so it works offline with no API dependency.

• Python
• ChromaDB
• MCP
• Embeddings

Zero-trust MCP server mirroring GitHub Copilot’s built-in VS Code tools, with strict policy enforcement and no cloud dependencies.

• Enforces deny/allow policy at the MCP execution layer rather than the prompt layer, so no amount of prompt injection or natural-language persuasion widens an agent’s access.
• Project policies layer on top of user policies and can only tighten, never loosen, with shell escape sequences blocked to stop nested commands from widening permissions.

• Go
• MCP
• Policy-as-Code
• Zero Trust

Transferred to Carleton. NCAA hockey.