AI Product · Productivity · Agentic

Case study by Shashank Srivastava

Jarvis: Designing and building a solo, agentic AI assistant.

An exploration of building a personalized AI assistant that manages emails, calendars, and productivity workflows, shifting from rigid keyword-based matching to a reasoning tool-caller.

01. Overview

Coordinating personal productivity tools often requires manual overhead. Jarvis is a solo, agentic AI assistant built to consolidate email, calendar tracking, and focus timers under a single conversational interface using LLM tool-calling.

Jarvis Case Study Cover

02. The Problem

Managing personal admin means constant app-switching. Email, calendars, notes, and productivity tools operate in silos without shared context, eating away valuable focus time:

The Problem: Bouncing between tools

03. The Vision

Creating a single conversational layer capable of parsing natural language, chaining operations, and executing actions across connected tools:

Vision: One conversational layer that can act

04. Target User

Designing for the high tech-comfort solo operator who acts as their own manager, operations lead, and administrator simultaneously:

Target User: Who Jarvis is built for

05. Approach & Phases

Detailing the end-to-end solo build timeline, moving from a foundational assistant loop to architecture refactoring and final reliability passes:

A solo build, shipped in phases

06. Key Decision

Pivoting mid-project from rigid keyword/intent matching to a dynamic LLM tool-calling engine to support messy, varied phrasing and multi-step requests:

Rebuilding the routing layer mid-project

07. Feature Modules

Reviewing the 10+ active modules registered as self-contained tools, including Gmail OAuth integrations, focus timers, and scroll blockers:

10+ modules, one interface

08. Interface Design

A chat-first, voice-ready conversational container featuring persistent session tracking and speech output:

A chat-first, voice-ready interface

09. User Journey

Illustrating how a single composite request is parsed and executed sequentially across multiple tools:

A single request, chained across two tools

10. System Architecture

The system flow mapping the interaction between natural language input, the core LLM orchestrator, and the tool-registry modules:

Architecture, at a glance

11. Engineering Challenges

Addressing reliability in tool selections, balancing LLM reasoning latency against user responsiveness, and sequencing solo project scope:

What was hard to get right

12. Project Reflection

Synthesizing key lessons in scoping, owning architecture trade-offs, and managing sunk costs when recognizing a design needs to be rebuilt:

What this project was really practicing

Read the next case study

Explore how I redesign group travel companion feature inside Google Maps navigation shell.

Read Google Maps Case Study