Case 05 · AI Personal Assistant

Say what you need.
Review what will happen.

Problem
Schedules, tasks and work results were scattered, and follow-up replies could lose the draft they referred to.1
Decision
Connected Mac, Telegram and mobile web, with separate steps for reviewing, approving, executing and checking a result.
Verified outcome
Checked real calendar reads, voice and text extraction from photos. Mobile approval and history recovery were verified with example data.
Personal projectMac + Telegram + mobile web
Concept diagram of requests, approval and result checks across computer and phone
Workflow concept · not an app screenshot or detailed implementation diagram.
My role
Problem definition · Conversation and execution boundaries · AI-assisted development direction · Integration and evaluation
Project name
AI Personal Assistant (Seoa)
Request from more places
Computer + phoneVoice, screens and messaging for everyday work
Before a change
Review → approve → actConnect intent to the exact change being proposed
After reconnecting
Check the saved resultInspect the existing record before starting again
Product experience

At the desk.
Away from it, too.

Use voice and a screen on Mac, or Telegram and mobile web elsewhere, to check schedules, tasks, goals and content production progress. Each channel keeps its own request and approval context.

01

From a result
to the next request

Check the day and ask a follow-up. Short replies stay associated with the candidates and change draft the user is looking at.

02

Speak, or read text
from a photo

Review and edit transcribed speech or extracted photo text before sending it. Listen to replies and stop playback when needed.

03

Review the change.
Then check the result.

Inspect a draft before changing a schedule or task. Save the outcome so the same request can be checked after reconnecting.

Design decisions

Reconnect without
doing the same job twice.

I chose to prevent duplicate changes and expose the actual state, even when that meant giving up the convenience of an automatic retry.

01

Keep the conversation
attached to its draft

Fixed cases where a short follow-up lost the actionable draft. Preserve the previous candidates and invalidate an earlier approval when the draft changes.

02

No response does not
mean nothing happened

A change may have been saved even when its response is lost. Do not repeat it automatically; show the uncertain state and check the existing record after reconnection.

03

Separate “nothing here”
from “could not check”

A failed lookup must not look like an empty calendar. Separate unnecessary lookups from the conversation and reduce repetitive diagnostic notices.

My contribution

Connect the conversation,
the screen and the action.

I defined useful tasks and failure criteria, divided AI-assisted development work, and reviewed the integrated result. AI supported implementation; product scope and final acceptance remained my responsibility.

Problem and criteria

Make the proposed change visible

Check that requested candidates are not lost, stale approvals cannot run again, and uncertain results never appear as success.

Technology and integration

Extend existing Mac capabilities to a phone

Connected Python-based work functions, conversation handling and saved records to the Mac app, Telegram and mobile web. Integrated voice and photo text extraction (OCR) into the reviewable input flow.

Evidence

Real connections and
example tests are different.

Everyday functions were checked through real tool connections. Approval and recovery exceptions were checked with non-personal examples and isolated tests.2

Actual connection and output checks
Mac calendar and reminder reads, real Telegram request/response handling, voice input/output, photo text extraction and delivery of a content-production notification.
Tests with non-personal example data
Mobile draft approval → result → reopening the saved record. Duplicate requests, restarts and a lost response after saving were checked in isolated tests.
Still to verify
Real mobile creation or editing of personal calendar entries, iPhone installation and microphone permission and push delivery, and long-term operation at a fixed address remain separate checks. The Mac and connected services must be available.
Earlier experiments are supporting history

Early functional tests and wake-word comparisons predate later changes. False wakes and missed real calls were evaluated together; those results are not reused as current product-wide performance.

What I bring

Beyond an answer,
a result you can check.

Understanding a request and changing external information are different responsibilities. I make that boundary visible in the product.

  • Conversation design
  • Approval and execution
  • Failure evaluation
  • AI development direction
Sources and scope
  1. Cross-checked the personal project’s current implementation, decisions, channel-specific flows and saved results. Mac, Telegram and mobile web reuse work capabilities; this does not imply that every conversation transfers seamlessly between channels.
  2. Actual connection checks are separate from isolated and synthetic-data validation. Real personal-calendar writes from mobile, unrestricted app control and always-listening mobile voice activation are not claimed as verified features. Private schedules, addresses and detailed operating rules remain undisclosed.