← Back to work

Case Study · 2026

AI Agent — One Assistant, Four Workflows

A Gemini-powered agent that drafts email, manages your calendar, and turns plain English into SQL.

FocusAI Agents · LLM Orchestration · RAG · NL2SQL
AI AGENT

Context

Everyday tasks — writing emails, scheduling meetings, querying databases — require switching between multiple tools. I built a single Gemini-powered agent that handles these workflows through natural language.

The Problem

People don't want to learn SQL, draft every email from scratch, or switch apps to check a calendar. The challenge was making one agent a reliable front door to several distinct systems.

Process

01

Tool orchestration

Designed the agent around specialized tool calls — email draft/send, calendar management, NL2SQL translation — each invoked based on detected intent.

02

Reliable NL2SQL

Built a pipeline that translates natural-language questions into valid SQL queries, with schema-awareness checks to reduce hallucinated columns/tables.

03

RAG for context

Added a retrieval layer so the agent's answers are grounded in actual documents and data rather than model memory alone.

Design Decisions

One agent, multiple tools — not multiple bots

A single conversational interface reduces cognitive load — the user states intent, the agent picks the right tool.

Outcome

✦ Automated email drafting and sending, cutting down composition time

✦ Natural-language database queries with no SQL knowledge required

✦ Responses grounded in real data via the RAG pipeline

Reflection

“The real challenge in agent-building isn't calling an LLM — it's knowing which tool to call, when, and how to verify the output is actually correct.”