I build the infrastructure that lets AI agents actually work inside enterprise software.

At Darwinbox I work on the agent platform — the layer between LLMs and real business systems. That has meant an Agent Builder framework in Python that lets engineers ship agents from declarative config, an MCP server exposing governed HR operations to external clients like Claude and Copilot, and native A2UI support so agent responses come back as interactive UI instead of plain text.

Most of the hard problems here aren't prompting. They're cost (multi-tier caching across prompt, tool definition and skill tiers), correctness (RAG grounding plus golden-set regression suites in CI/CD), and trust (every action permission-checked and traceable).

I also own the Angular frontend for Darwinbox Super Agent and built a shared Stencil Web Components design system used across product teams. Working on both the runtime and the interface is the part I like most.

Agent runtimes — tool registries, declarative agent config, execution boundaries, and keeping LLM reasoning inside the permission model it was given.

Making inference affordable — caching at the prompt, tool-definition and skill layers, because token cost is the thing that decides whether an agent feature ships or dies in review.

Evaluation — golden-set regression suites wired into CI/CD, so an agent can't silently get worse between deploys.

Interfaces for agents — A2UI, streaming responses, and framework-agnostic components. Chat is not the only shape an agent can take.

• DeskDoc — AI voice receptionist for clinics. Vapi for call orchestration, OpenAI for intent handling, Node.js backend with webhook-driven call flows and real-time slot management.
• FinovaAI — personal finance app with a financial health scoring engine, Redis caching, and push notifications.

Conversations about agent infrastructure, MCP and A2UI, evaluation harnesses, and anything at the intersection of AI systems and product engineering.