Selected work

AI engineering / 2026

Wixal

A local AI workspace with reviewed tools, persistent conversations and a real terminal.

My role
Product design and application development
Stage
Local macOS build
Built with
Electron · JavaScript · Ollama · node-pty · xterm.js
Wixal macOS interface with local model selection and project workspace
Actual application interface.

01 / Context

The problem

An AI coding workspace needs more than a chat box. It needs project context, usable tools, durable state and a clear point at which the user decides what can run.

02 / Ownership

My contribution

I built the desktop interface, local Ollama integration, project and conversation persistence, tool review workflow, and terminal integration.

01Project context
02Local Ollama model
03Structured tool request
04Human review
05File or command result

03 / Reasoning

The decisions behind it

01

Keep inference local

The app talks to Ollama on loopback. Model selection and streaming sit alongside project context, with no cloud provider configured.

02

Review before action

File changes show existing and proposed contents. Model-requested commands require review. The terminal and approved commands still run with the user’s host permissions.

03

Build a desktop application

A narrow preload bridge connects the renderer to the main process. Persistent state is written atomically, and native terminal modules are packaged with the application.

04 / In practice

A closer look

05 / Evidence

Verification & boundaries

The project includes unit checks and an app smoke workflow covering a PTY, reviewed fixture-file creation, explicit memory, recall and reload. The screenshots shown here come from that controlled workflow; they do not establish reliability for every model.

What this does and does not establish

The build is for Apple Silicon and is not signed or notarized for public distribution. Local JSON storage is not encrypted. Reviewed commands are host shells, not a filesystem sandbox.

06 / Looking ahead

What comes next

Improve evaluation across models and failure cases, refine file diffs, and prepare signing and distribution before offering a public download.

Next case studySentinel Local

The next chapter

Good work starts
with a conversation.

Engineering opportunities, thoughtful teams, and useful problems in software, AI and cybersecurity.

Get in touch