I build AI infrastructure in TypeScript and Python: multi-provider SDKs, durable agent runtimes, and MCP integrations. My background in banking analytics shapes how I approach data quality, traceability, and reliable systems.
I contribute to the Zhivex ecosystem, with public work on agent safety, persistence, provider integrations, and release validation.
Build agents and AI applications around a shared model contract, with explicit provider capabilities, tools, streaming, and durable execution.
My contribution: hardened retries, routing fallbacks, and agent generation safety; added strict unknown-tool validation and recovery. Contribution · Quickstart · Human approval example · npm package
Build async agents with tools and a portable provider contract, then add durable state, approvals, and observability as the application needs them.
My contribution: implemented durable onboarding and operational evidence, including preservation of typed tool inputs when resuming after approval. Contribution · Quickstart · Runnable agent · PyPI package
Connect agent tools to official APIs from Argentina's Central Bank. For example, discover a public statistical series and retrieve its history with validated inputs and structured responses.
My contribution: expanded BCRA API coverage and refactored the tools and shared HTTP infrastructure. Contribution · Setup and tools
I build and maintain Power BI dashboards for banking teams, combining semantic modeling, DAX, SQL, Power Query, data quality checks, and operational reporting. My work turns reporting requirements into defined KPIs and data models that support executive and operational analysis.
Banking projects and their semantic models remain confidential. Public examples are independent demonstrations created with synthetic data.
Explore the synthetic analytics example — dataset, SQL, a visual summary, and a Power BI modeling recipe.
Run the approval demo: an agent proposes a simulated action, pauses for approval, and executes it only when you approve. It uses the real Zhivex agent runtime with a scripted model, synthetic inputs, and no API keys. You can also deny the action and verify that it never executes.
Based in Buenos Aires, Argentina. I work in English and Spanish.
For conversations about AI infrastructure, SDK integrations, or data analytics, connect with me on LinkedIn or X.