03 / 11Flagship / Local-First AI Assistant

MeBot

A local-first personal assistant for capturing unstructured thoughts, turning them into actions, and managing the workspace through an agent interface.

RoleProduct architect & full-stack engineer

Period2026

StatusActive personal productivity product

EvidenceLive authenticated product, privacy-safe captures, and a sanitized public engineering case study

Visit live project View repository

01Context

The engineering problem

Challenge

Make fast, messy capture useful without forcing people to organize every thought up front or making private notes depend on a constant network connection.

Response

Designed an IndexedDB-first experience with an offline write queue, structured AI triage, semantic memory, a client-side vault, and an agent mode for changing workspace behavior through natural language.

Outcome

A deployed PWA that demonstrates privacy-conscious product boundaries, resilient local interaction, compound-input handling, and integrations with Google Calendar and Tasks.

02System architecture

How the layers connect

The diagram is paired with a text alternative and remains readable without animation or WebGL.

01CaptureA low-friction dump interface accepts unstructured and compound input.

02Local stateIndexedDB and an offline queue keep the primary interaction resilient.

03AI triageStructured output turns raw input into notes, tasks, categories, and follow-up actions.

04MemoryVector search and controlled context retrieval support continuity across sessions.

05IntegrationsCloud services connect reminders with Google Calendar and Tasks.

Capture: A low-friction dump interface accepts unstructured and compound input.
Local state: IndexedDB and an offline queue keep the primary interaction resilient.
AI triage: Structured output turns raw input into notes, tasks, categories, and follow-up actions.
Memory: Vector search and controlled context retrieval support continuity across sessions.
Integrations: Cloud services connect reminders with Google Calendar and Tasks.

03Engineering decisions

Choices that shaped the result

Keep capture local-first

The interface writes locally before synchronization, so the most frequent action remains fast and available when connectivity is unreliable.

Separate dumping from agent control

Dump mode is optimized for capture; Agent mode is reserved for deliberate workspace changes such as categories, themes, reminders, and memory rules.

Publish the system, not private data

The public case study documents architecture and interaction patterns while excluding notes, database contents, credentials, vault data, and security-sensitive implementation detail.

04Product evidence

Selected interface views

Captured from the deployed product. Authenticated screens are shown only in privacy-safe states.

MeBot agent workspace configuration interface — Agent mode / natural-language control for workspace configuration

05Technical footprint

Stack in context

Next.js 16
React 19
TypeScript
IndexedDB
Cloudflare Workers
D1
R2
Vectorize
Gemini

Local-first capture
Offline write queue
Semantic memory & agent mode

Next case studyHMM ITB Platform