Gemini Agentic Tool-Use: 75+ MCP Service Simulations & ADK Plugins

75+Stateful MCP Service Simulations-88%Payload & -58% Schema Tokens+7.7ppAccuracy (~99% Solvability)

System Architecture Flow

01Ergonomic MCP Toolkits (-88% Payload / -58% Tokens)

0275+ Stateful Simulations (Pydantic v2 + SQLModel)

03Google ADK Multi-Service Plugins & Gradio UI

04Containerized Multi-Turn Benchmarks (Docker + Pytest)

01 / Engineering & Implementation

Collaborated with Google to train and evaluate Gemini tool-use by architecting high-fidelity, stateful API simulations across 75+ services (Gmail, Slack, Jira, Shopify, Calendar, GitHub) using Python, Pydantic v2, and SQLModel/SQLite.
Engineered an ergonomic MCP toolkit and schema-minimization layer for Gemini that flattened deeply nested schemas, reducing response payload size by ~88% and schema token overhead by ~58% while lifting accuracy by +7.7pp (~99% task solvability).
Built cross-service agent plugins on Google ADK with a local skill registry and interactive Gradio streaming UI to orchestrate multi-step workflows across workspace, commerce, and payment connectors.
Developed containerized multi-turn coding benchmarks (Docker + Pytest) and collaborated with Meta on complex Python codebase reasoning, cutting per-service engineering delivery from days to hours with reusable AI skills.

~88% smaller payloads, ~58% fewer schema tokens, +7.7pp accuracy gain, and ~99% solvability across 75+ services