Local inference
A native SwiftUI interface and MLX runtime with streamed responses. Model selection also takes input type and availability into account.
01 / iOS · Local AI
HomeHubPersonal AI. Closer to you.
In HomeHub, I bring a personal AI assistant together with my documents and conversations. It runs on iPhone and iPad, and I decide what it remembers.
TECHNICAL CURIOSITY + PERSONAL USE
How much of an assistant can fit on a phone?
I wanted to ask questions about my own documents on a phone, check the source behind an answer and decide what the assistant remembers. HomeHub is my personal project exploring how much of this work models can do on the device.
The chat window is only the beginning. The project addresses text and vision model selection, user-approved memory, source retrieval, limited memory and the mobile app lifecycle.
- 01
Add a source
A document enters the knowledge base. OCR can help with scanned PDFs.
- 02
Find an answer
Relevant excerpts provide context and a basis for citing the source.
- 03
Decide what to remember
Useful information can become a suggestion that the user accepts or rejects.
BROWSER SIMULATION
Your assistant. Right on your phone.
Try selected workflows directly in your browser.
This is a separate simulation using fictional data.
HOMEHUB / INTERACTIVE DEMOSIMULATION · NO SIGN-IN
Loading the demo… A text overview is available below.
What you can try in this demo
- Home — Local model status, quick actions and recent conversations.
- Document and citation — Open a fictional source document, display the prepared answer and check it against the cited excerpt.
- Models — Filter the sample catalogue and choose the active model.
- Memory — Accept a suggested memory and browse the documents.
- Background behaviour — In Settings, select App in the background to see why the model is unloaded from memory when you leave the app. Select Return the app to the foreground to reset this prepared example.
TECHNICAL DECISIONS
An assistant is more than a single model call.
Your own sources
PDF and text import, OCR for scanned pages, and retrieval of relevant excerpts to support citations.
Memory under your control
The assistant can suggest information to remember. The user decides whether to accept, reject or remove it.
Choose a model for the task
Selection takes image input, the user's pinned choice and the model's reliability record into account.
Memory has its own lifecycle
Suggesting information and approving it are separate steps. Being able to remove it later matters just as much.
Mobile constraints shape the design
Background transitions and available memory headroom affect how the resident model is managed.
My contribution
Design and development of my own iOS project, bringing conversation, the model runtime, memory and native services together in one interface.
Project and demo status
This personal iOS project connects on-device AI, documents and user-controlled memory. The website demonstrates these ideas with prepared responses. The native app and its device performance were not tested again while preparing this portfolio.
EVIDENCE AND LIMITS
The description is based on local code, test definitions and supplied screenshots. Local development is ahead of the verified public main branch. Local inference does not mean that every feature works without a network: search and model downloads may use one.