From the last result
to the next question
One call opens a conversation. The user can ask more about a place or result already shown, or revise the draft currently being discussed.
Less effort spent repeating the same context
Case 05 · Seoa / Athena
Keep talking.
Move the day forward.
A personal Mac assistant connecting schedules, tasks, and research through voice and screen. I designed the experience and coordinated five AI development workstreams into one application.1
Personal projectIntegrated and being refined
- My role
- Problem definition · Conversation design · AI development direction · Integration and validation
- Collaboration
- AI implemented assigned work; I set the scope, priorities, and acceptance criteria
- Connected to existing information
- iPhone ↔ MacWorks with calendars and reminders already in use
- Spoken and visible
- Voice + screenAsk briefly; review the result as you go
- One integrated product
- Personal Mac appSeparate capabilities, one user experience
Product experienceWithout explaining
Without explaining
everything again.
From the initial call to choosing a next action, I focused on keeping the conversation intact between features.
From the last result
Brief speech.
Visible detail.
Instead of reading an entire schedule or draft aloud, Seoa gives a short response and shows the result alongside it. The user can inspect the details while listening.
The convenience of voice, with the clarity of a screen
From information
to a next step
Review the day's schedule and tasks, then research what is needed. A selected result can lead into a proposed calendar or task change.
Connect viewing, research, and changes in one place
Design decisionsA natural conversation.
A natural conversation.
Deliberate decisions.
Alongside the features, I defined what the user needs to understand and decide at each step.
The user's context,
not an isolated sentence
Interpreting every short follow-up as a separate command broke the conversation. I shifted the experience toward considering the previous result and pending draft, while showing what the assistant understood.
“Nothing there” is different
from “could not check”
A failed lookup should not look like an empty result. I made retrieval failures and synchronization state visible so the user could judge the information correctly.
Flexible interpretation.
Explicit changes.
AI interprets natural language, while meaningful external changes pass through a draft and confirmation. The user should understand what will change before proceeding.
Development directionSeparate implementations.
Separate implementations.
One product.
Dividing the work does not finish the product. I coordinated shared interfaces and integration points, then checked the combined result against one standard.
Phosynd · Decisions and coordination
Problem, scope, and acceptance
I selected the assumptions to test first, assigned ownership, and set the integration order. The user experience and quality of the complete product remained my responsibility.
AI · Assigned development work
Implementation and revision
Separate development sessions implemented their assigned capabilities and incorporated review findings. Their outputs were reviewed, revised, and connected into the complete application.
- Define the problem
- Assign responsibilities
- Integrate
- Validate the whole
EvidenceCheck the function.
Check the function.
Refine the experience.
Automated tests checked functional connections, and device checks covered everyday interactions. Wake-word behavior was refined through a separate recording comparison.2
- Automated functional tests 476 passed
- A saved quality-validation record covering calendar, task, conversation, and confirmation behavior. These tests isolate live model calls and personal-data changes.
- Device checks Daily flows
- The acceptance record marks briefing, voice-approved calendar updates, speech interruption, and phone-entered tasks and priorities appearing in the app as completed.
- Wake-word tuning experiment Tradeoff reviewed
- On the same recordings, false activations fell from 32/63 to 0/63, while accepted real calls also fell from 139/192 to 125/192. I compared the reduction in false activations with the increase in missed calls.
Scope of the recording comparison
The saved experiment used the same sets of 63 noise and non-command recordings and 192 real-call recordings for the comparison. These observations apply to those recording sets, not overall app accuracy or freedom from false activations in everyday use.
What I bring
Turn AI-built parts
into a product people can use.
With Seoa, I defined the problem, connected separate implementations, and validated the result against clear acceptance criteria. The work combines implementation with the judgment needed to make one coherent experience.
- Product definition
- Conversation UX
- AI development direction
- Integration and validation
Sources and scope
- Personal project design, development coordination, and acceptance records. Five refers to development workstreams covering implementation and separate review, not assistants running inside the app. Implemented calendar, task, and research functionality is distinguished from completed user acceptance checks.
- Saved quality documentation, automated test definitions, and wake-word comparison records. Automated functional tests, device-level acceptance checks, and recording comparisons are separate validation measures. Each result is scoped to functional behavior, the user flows checked, or the specific recording sets.