I use JaJ to find information, research things and help manage my day. Browsing the web has become a frustrating part of finding answers, and I wanted an assistant that could help with that work while keeping my own information within reach.
The biggest benefit for me is time saved and better organisation. JaJ is designed around how I work, so its notes, tools and workspace can fit together in ways I don’t often find in other applications.
I also care about how it feels to use. The interface should help me understand what is happening and keep relevant material available as the assistant works. I find that satisfying, and it makes me want to return to the application.
Giving the model somewhere to work
JaJ connects a local AI model to my notes and a defined set of tools. The model can request operations, receive their results and use that information in the next part of the conversation.
For note search, those tools find and rank relevant material, then read it back into context. Searching by words, and by meaning when available, gives the model material from my own files to work with. I can return to information without having to remember exactly where I put it.
Sources matter too. JaJ keeps an evidence record and checks that citations point to material its tools actually returned. I can open those sources and make my own judgment about an answer. A reference I can inspect is part of the workflow, rather than decoration at the end of a response.
The workspace connects conversation to the rest of the application. Notes, tasks, sources and a calendar can be brought into view through defined operations. That is where the assistant starts helping me organise the work around a question, as well as helping answer it.
Making the pieces dependable
JaJ is one of my most complex projects. A rudimentary version of each piece is fairly straightforward. Refining the behaviour and making those pieces work reliably together has taken most of the effort.
Custom note search, context management, questions, tool use and the information appearing on screen all affect one another. The model needs enough context to continue the task, while the application needs to keep control of the operations it can perform and the outcomes it reports.
A reviewed note change is a concrete example. JaJ prepares an exact preview and waits for my approval. The native application checks the file’s revision and permissions before saving. A refused or stale proposal must remain a proposal; it cannot become a claim that the file was changed.
I also create and refine AI harness workflows and skills for my own needs. The harness is the software and rules around the model that connect it to a job. Choosing tools, handling context and checking results is a substantial part of making AI fit a workflow.
A project shaped by daily use
JaJ remains a personal project in development. Using it helps me see which integrations save time, which information should stay available and where the behaviour needs more refinement.
Getting those details right has been a challenge I’ve enjoyed. Seeing the application emerge, and having it help me stay on top of things, is what makes the work worthwhile.