Case Study

We turned our founder’s WhatsApp into an interface for running AI Garage

How we built Shushi — a personal AI agent that works inside our founder’s existing WhatsApp conversations to understand requests, check context, use business tools, ask for approval, and take action.

Built for

AI Garage — internal dogfooding

Status

Live · used by AI Garage’s founder

Interface

WhatsApp

S

Shushi

AI Agent · WhatsApp

Shushi, can you book an intro meet with Vedant for coming Monday at 2 pm for 1 hr?

v*****@gmail.com

10:42 AM

Sure! I’ve set up the details for your meeting with Vedant:

Intro meet with Vedant

Monday · 2:00–3:00 PM IST

v*****@gmail.com

Company calendar

Shall I go ahead and book this for you?

10:42 AM

yes pls

10:42 AM

Done! I’ve booked the intro meet with Vedant for tomorrow at 2:00 PM IST. Invitations have been sent.

10:42 AM

16 seconds from request → booked.

The State Before Shushi

The problem wasn’t Google Calendar. The meeting request was already happening inside WhatsApp. But completing it meant leaving the conversation, opening Calendar, checking multiple calendars, finding conflicts, creating an event, copying details, adding attendees, configuring the meeting, sending the invite, and returning to WhatsApp.

The same fragmentation appears across dozens of founder workflows. So instead of building another dashboard, we asked:

What if WhatsApp itself could become the interface to AI Garage’s operations?

The Challenge

A useful WhatsApp agent couldn’t simply translate messages into API calls. It needed to understand incomplete requests, remember what was being discussed, distinguish between conversations, search multiple calendars, respect Shubham’s preferences, detect conflicts, ask for missing information, and know when not to act.

Most importantly, consequential actions needed to remain under human control.

The goal: make operating AI Garage conversational without making it unpredictable.

Real Usage

Since we started recording Shushi’s activity:

196

Messages processed

96

Shushi responses

23

Active usage days

23

Completed calendar actions

~1+ hr

Direct calendar administration automated

Conservative estimate of 3 minutes of manual calendar administration per completed action. Excludes availability checks, context switching, memory retrieval, and other Shushi workflows.

One Conversation. Multiple Systems.

A request as simple as:

“Shushi, book calendar with Ronit for tomorrow 2pm”

can trigger an entire workflow.

01

Understand

Extract the person, date, time, and intent from natural language.

02

Check

Search relevant calendars and operating rules.

03

Clarify

Ask for duration, email, or calendar when required.

04

Remember

Carry those details across subsequent WhatsApp messages.

05

Confirm

Summarise the proposed action before doing anything consequential.

06

Act

Create or update the event and send invitations.

07

Continue

Keep the resulting action available as conversational context.

It Doesn’t Just Follow Commands

Shushi understands the operational rules around every action.

It knows operating preferences

When a Sunday meeting was requested for 1 PM, Shushi found an existing Fellowship session and knew Shubham’s preferred Sunday booking window started at 2 PM. Instead of blindly booking, it proposed the nearest appropriate slot.

It knows when information is missing

If a meeting request doesn’t contain a duration or attendee email, Shushi asks rather than inventing one.

Preferences aren’t hard restrictions

If Shubham explicitly wants a meeting outside his normal booking window, he can override the preference conversationally.

It works across identities

Shushi can reason across Shubham’s Personal, Company, and AI Garage calendars and use the appropriate one for the task.

Review Over Autonomy

We deliberately didn’t design Shushi to act autonomously whenever it thinks it understands something. Before consequential actions, it shows what it intends to do and waits for confirmation.

That also creates a recovery path when AI gets something wrong.

Real conversation — error correction

Meeting seems to have been scheduled till 8 pm?

I’ve prepared an update to change the end time to 2:30 PM IST… Shall I go ahead and fix this for you?

Yes pls.

The goal isn’t zero mistakes. It’s controlled, inspectable automation.

Architecture

The complexity stays behind WhatsApp.

WhatsApp
Shushi
Context
Memory
Rules
Tools
Google Calendar + AI Garage systems

Under the hood

WhatsApp / Baileys

The conversational interface and message transport.

n8n

Self-hosted orchestration for deterministic workflows and tool execution.

Gemini

Natural-language understanding and reasoning.

Supabase

Conversation state, pending actions, and operational context.

Google Calendar

Availability, event creation, and event updates.

Custom APIs

The layer that lets Shushi progressively reach more AI Garage systems.

What Makes It More Than a WhatsApp Bot

A chatbot returns an answer. Shushi can complete an operation.

Messageunderstands intentretrieves contextchecks business rulescalls toolsasks when uncertainwaits for approvalperforms the actionpreserves state

That distinction changes WhatsApp from a messaging surface into an operating interface.

Beyond Calendar

Calendar was the first production workflow, not the end state. Shushi is already capable of retrieving information from Shubham’s broader context.

Real conversation — AI Garage WhatsApp group

Shushi, what was my review of profound profile?

Shushi retrieved his previous observations and returned the relevant context directly into the group conversation.

The interface hadn’t changed. The capability underneath it had.

Why we built it for ourselves first

Shushi wasn’t built as a portfolio demo. Shubham actually uses it while running AI Garage. Every awkward clarification, failed action, missing integration, and edge case becomes something we experience ourselves.

Dogfood the automation before recommending the pattern to someone else.

Results

23 completed calendar actions executed through the agent.

Multiple calendars → one interface Company, AI Garage, and personal context can be reasoned about from WhatsApp.

Human approval built in Consequential actions wait for explicit confirmation.

No new interface to learn The agent lives inside a tool already used every day.

Request → booked in ~16 seconds A real meeting request has gone from WhatsApp message to confirmed calendar booking.

Self-hosted orchestration The automation layer runs on infrastructure controlled by AI Garage.

Shushi Is the First Agent, Not the Last

The interesting part isn’t building a calendar assistant called Shushi. It’s the underlying pattern:

Person + existing interface + company context + business rules + tools + human approval

Change those components and the same architecture can support very different jobs. We’re already exploring this pattern with assistants such as Loop for AI Garage’s community operations, and extending Shushi itself to interact with browsers — starting with the ability to fill any form shared via WhatsApp using the user’s existing context, with human approval before submission.

For another business, the interface could be WhatsApp, Slack, email, or something internal. The workflows could be sales, hiring, support, finance, or operations.

What would your Shushi do?

Think about the work that already begins in your team’s WhatsApp messages, emails, spreadsheets, and internal conversations. What happens next manually? That’s where we start.