Flora Mail BETAGitHub (may require access, opens in new tab)
Flora Mail – Beta v0.2.0 Prerelease

Flora Mail.

Intelligence, in its place.

A local-first, AI-augmented email client exploring how intelligence should live inside traditional graphical applications.

Explore the project View research
Familiar structure.
Thoughtfully augmented.
01 / THE FLORA HYBRID EXPERIENCEConceptual presentation · Fictional Beta demo content
Native macOS applicationProduct design · Software engineering · HCI / AI research

01 The design question

More intelligence.
Less interface?

AI opens up new ways to work. But when everything becomes a conversation, what happens to the structure we rely on?

01 / StructureChat-first interaction can hide navigation and make familiar tasks harder to see.

02 / ControlUnpredictable actions can leave users unsure of what changed, or why.

03 / PrivacyCloud-first intelligence can expose personal context beyond the device.

How should artificial intelligence be integrated into traditional graphical applications?

02 Design philosophy

The interface stays. Intelligence joins.

GUI provides structure and control. AI adds understanding, organization, prediction, and suggestions. The assistant is an optional way to access information.

GUIStructure & control
+
AIUnderstanding & assistance
=
FLORA HYBRIDA familiar interface.
A more capable you.

GUI completeness

Every essential task has a graphical path.

AI assistance

Understanding, organization, and suggestions.

Human control

Review and confirm meaningful actions.

Reversibility

Visible history and clearly bounded Undo.

Designed for predictability, adaptability, transparency, progressive automation, and local privacy.

03 Product experience

Email first. Assistance where it helps.

A native SwiftUI workspace, from the first message of the day to the last draft. Demo Mode is available without accounts or a model.

Inbox · Reader · Categories

Keep the original within reach.

Source-backed summaries, classification suggestions, and extracted actions sit inside a conventional message reader. Accept a local category when it makes sense. Keep reading when it doesn’t.

Smart Reply offers editable ideas in the normal composer.

Message Reader
PROFESSOR ELAINE LIN · FICTIONAL MAIL

PHYS 204 · A new deadline

At a glance

Submit the problem set by Friday, 5:00 PM.

Suggested categorySchool
Review suggestion → Confirm category

Conceptual presentation · Not an application screenshot

04 Local intelligence

Private
by design.

Personal mail deserves a deliberate boundary. In Local AI mode, inference stays on the device, with no silent cloud fallback.

Limited contextOnly the context needed for the request.

Validated outputStructured responses and source-quote checks.

Cancellable requestsLate results cannot overwrite a different task.

Local AI cachingCapability caches respect content and account boundaries.

ON YOUR MAC
  1. 01
    Mail

    Selected, locally cached messages

  2. 02
    Local context

    Bounded, task-specific excerpts

  3. 03
    Local AI provider

    Validation, cancellation & cache

  4. 04
    Ollama

    Loopback runtime · Local inference

  5. 05
    gemma3:4b

    Verified on-device model

Mail sync separately contacts your authenticated provider. Local AI stays inside the local inference boundary.

DEMO AI / LOCAL AI / OFF

AI mode is an explicit choice. Generated text can be wrong; source validation does not guarantee accuracy. Local mail storage is a prototype JSON store, not an encrypted database.

05 Human-guided AI

The model suggests. You decide.

Suggest → Confirm → Execute. AI providers hold no mail writer or action executor. Permission and execution belong to the application’s controlled action layer.

  1. 01

    AI understands

    Context, meaning, and possible next steps.

  2. 02

    AI suggests

    A visible, editable proposal.

  3. 03

    User reviews

    Check the source, scope, and effect.

  4. 04

    User confirms

    Explicit approval for the action.

  5. 05

    Application executes

    Permission checks and a durable record.

An architectural boundary, by intention. The assistant is read-only. Meaningful external actions require user-defined authority and application checks. Remote sends do not offer a fictional Undo.

06 Progressive automation

More assistance. Deliberate authority.

Automation can increase. The scope of permission is always defined by the user. Explore the five levels implemented in the research prototype.

More manual interactionMore explicitly scoped assistance

Level 2 / Suggestion

A useful suggestion. An explicit decision.

AI can suggest a category or prepare a writing preview. You review and accept the suggestion. External mail actions still pass through account permissions and a confirmation preview.

Permission boundary

Each suggestion requires explicit acceptance.

Accept a category or review an editable reply.

07 Architecture & engineering

Built close to the platform.

A native macOS application built with Swift and SwiftUI. Clear boundaries separate presentation, intelligence, permissions, and the systems that act.

PresentationSwiftUINative macOS interface
App / Interaction LayerNavigation · State · Review & confirmation
DomainIntelligenceAutomation
Mail ProviderGmail · Microsoft · iCloud
AI ProviderDemo · Local AI · Off
Local storageKeychainNetworkOllama

Simplified from the Beta architecture. The reasoning layer has no direct mail-write authority.

Documented v0.2.0 verification422 / 422

Automated tests passed.

130 XCTest + 292 Swift Testing.
Debug and Release builds passing.

Engineering evidence from fictional mail and provider fixtures. Live-account validation remains unverified.

Swift 6 + SwiftUINative macOS research prototype

Swift Package ManagerExplicit modules and provider abstractions

Local-first persistenceOffline mail cache · Keychain credentials

3 mail provider architecturesGmail · Microsoft Outlook / 365 · iCloud

What the prerelease does—and does not—establish

The adapters include authentication, bounded synchronization, and many confirmed write operations. Live Gmail, Microsoft, and iCloud sign-in, synchronization, and writes remain unverified; real-account validation is ongoing.

The local preview is ad hoc signed. Production readiness, notarized distribution, comprehensive native accessibility coverage, and broad hardware validation are not claimed. The portfolio’s browser checks are separate from the macOS application’s documented 422-test gate.

08 The research behind the product

Working paper · Research prototype
HUMAN–COMPUTER INTERACTION / APPLIED AI

Designing for
Intelligence.

A Human-Guided Framework
for AI-Augmented Applications

Flora Mail — primary case study

How can an application become more intelligent without making its user less in control?

The proposed framework treats AI as an augmentation of an already complete graphical application. Structure stays visible, suggestions stay reviewable, and authority stays explicit.

Flora Mail puts that framework into a working macOS email prototype: a familiar GUI, embedded assistance, and an optional information-access assistant.

Implementation is evidence of a design direction. It is not evidence of a usability improvement.

Three lenses on the same task.

Implemented research presentation modes
A

Traditional GUI

Ordinary mail navigation and controls. AI and contextual assistance are hidden.

GUI
B

Chat-First

The read-only assistant leads the Today surface. Normal mail controls remain accessible.

AssistantGUI
CPROPOSED DIRECTION

Flora Hybrid

A traditional GUI with embedded AI and an optional assistant. The default research mode.

GUI+Embedded AI+Assistant

These modes change presentation, not permissions. They do not establish feature parity or a controlled experimental design.

Proposed Evaluation

Design the study
as carefully as the interface.

A future evaluation could compare how people find information, review suggestions, and complete email tasks across the three conditions.

No participant study or experimental results are reported.
01 / Tasks & comparisons

Proposed tasks include finding a message, identifying a deadline, organizing mail, and drafting a reply. Tasks must be reviewed for suitability in each mode before any comparison. Assignment and counterbalancing remain to be designed.

02 / Measures & human control

A proposed study could examine task completion and errors, review and correction behavior, perceived control, predictability, trust, and workload. Task correctness and subjective measures need a reviewed protocol; they are not inferred from event counts.

03 / Consent & privacy

Existing instrumentation is opt-in and off by default. It records fixed interaction events and explicit task timers locally, excluding mail content, account identity, queries, and credentials. A real study still needs recruitment, participant consent, a data-handling plan, and any required institutional review.

04 / Current evidence & limits

The prototype and automated developer checks establish implementation behavior. They do not establish statistical significance, usability gains, model accuracy, or experimental findings. The working paper’s publication status and public manuscript are not yet provided.

09 Project journey

From a question to a working prototype.

  1. 01Concept
  2. 02Interaction framework
  3. 03Native macOS prototype
  4. 04Local AI
  5. 05Mail providers
  6. 06Real mail actions
  7. 07Context intelligence
  8. 08Progressive automation
  9. 09Research prototype

A conceptual development sequence. Provider-backed actions are implemented; live-account validation is still ahead.

Research Prototype

Built. Tested.
Still being explored.

Flora Mail – Beta / v0.2.0 Prerelease

The idea worth pursuing

Designing AI that works
with the interface
— not instead of it.

Explore on GitHub (may require access, opens in new tab)Research overview
Repository may require access.Paper / Public manuscript to be addedContact / Details to be added