Is anyone into AI? What are you using? Claude Pro? Gemini 2.5? Some amazing applications..

[computerwiz]

15 hours ago, hybridxer0 said:

Sounds like you have a lot of tinkering like me! I used to own an IT firm before I sold it and moved, but I always will be tech-adjacent in whatever I do. Let me know if I can help in any way!  

 

If you find any good offline, portable, tools, I'm all ears.  Things move so fast in the AI world, and there's so many flavors, it helps to get suggestions from others.  Here's what I'm settled on (so far):

KoboldCpp - General all-purpose AI engine.  Good for chats, TTS, Image analysis, simple image gen, and eventually music generation.  Now has video generation ability (though my PCs are too weak to even try).

 

AnythingLLM - Extends the abilities of KoboldCpp, with memories & file reading ability.

 

ComfyUI - Any kind of image generating or editing.

 

GGUFLoader - Similar to KoboldCpp, but much simpler.

 

That's what I have so far.  And I'm taking a break working on it for now.  But definitely feel free to contribute any findings you have that fit what I'm looking for: free, offline, portable, & CPU-friendly.  

 

8 hours ago, Desert Dan said:

 

I have some general chat type models running local on just 8gb of VRAM, including image analysis, etc, but I have yet to successfully figure out text to image generation running locally.  ComfyUI made my brain hurt and it always seems like packages or components are outdated, erroring out, missing, etc.  I have to step back from things like that, or else I will get frustrated.  I like messing with microcontrollers and junk like that as well.

 

I don't know how long ago you messed with ComfyUI, but they have templates now.  Also it may help to have the ComfyUI Manager installed.  All you need is a good image gen model, and a good positive & negative prompts.  Some fine tuning like the # of steps & the CFG value, but that's about it...

[Paulus]

Very interesting and has a lot of potential.

 

[Cool Cat]

I thought it might be interesting to give some concrete examples of how I use AI, because my use of it has moved quite a long way beyond simply opening ChatGPT and asking questions.

 

Yes this has been written by AI but thoroughly checked and heavily prompted to get there. 😀

 

One of my main experiments has been building a self-hosted personal AI agent called Bud.

 

Rather than being just another chatbot, Bud runs continuously on my own infrastructure and has gradually become a kind of personal AI operating layer.

What Bud actually does

Spiritual study

Bud helps me with quite a lot of spiritual study and preparation.

It can:

• prepare the day's text with key principles, meditation points and possible personal applications
• produce useful comments on every verse of scheduled Bible reading
• conduct deeper chapter studies using an indexed local library of publications
• search jw.org/WOL material that I have indexed locally
• use verified Bible wording rather than inventing or loosely paraphrasing scriptures

I've built a fairly substantial local spiritual research library behind it.

That now includes:

• more than 105,000 indexed passages from books, magazines, meeting workbooks and other material
• the complete NWT Study Edition — 1,189 chapters, including study notes, footnotes and cross-references
• around 2,260 Watchtower study articles
• 19 books cached from WOL

So instead of relying entirely on general AI training data, the agent can search the actual source material before answering.

That has been one of the most useful things I've learnt about AI:

Give the AI good source material rather than expecting it to know everything.

---

Health, fitness and rehabilitation

I've also built a fully integrated fitness and rehabilitation application.

It records workouts, exercises, progress and rehabilitation information, synchronises the information with my Obsidian vault, and then makes relevant information available to Bud.

That means my AI agent can have an ongoing picture of my training and rehabilitation rather than each conversation beginning from scratch.

It can help me:

• keep track of exercises and progress
• compare training sessions
• record rehabilitation information
• follow changes over time
• connect workout information with my wider notes

I don't see this as replacing medical or professional advice.

It's more like having a very organised record system that can also reason over the information already stored in it.

---

Organisation and administration

Bud also handles a lot of the everyday organisational work that is easy to forget.

It can:

• maintain my calendar
• manage reminders and appointments
• keep a dedicated scheduling conversation as a single source of truth
• maintain durable notes
• record things I don't want to forget
• keep information in an Obsidian knowledge vault rather than losing everything when a chat ends

This persistent information is something I find particularly useful.

A normal chatbot conversation can be temporary.

My aim has been to create something where useful information gradually becomes part of a structured personal knowledge system.

---

AI helping with my secretary work

Another very practical use I make of AI is writing and administration.

In my congregation secretary role there are quite a lot of occasions where information needs to be communicated clearly, accurately and succinctly.

AI is extremely useful for taking what I already want to say and helping me turn it into something clearer.

For example, I use it to help with:

• organising rough notes
• making messages easier to understand
• shortening something without losing important information
• preparing concise administrative messages
• making wording warm and tactful
• removing repetition
• turning a collection of points into a logical message

That is actually one of my most frequent uses of AI.

Sometimes I know exactly what I want to communicate, but the AI can help me express it in two minutes rather than spending twenty minutes trying to get the wording right.

For me, AI is often less about writing something for me and more about helping me communicate my own thoughts clearly and efficiently.

---

Daily briefings

Every morning at around 6:15am, Bud prepares a briefing covering things such as:

• what's on the calendar
• important reminders
• current priorities
• relevant health information
• things approaching that I need to think about

It can also automatically produce:

• tomorrow previews
• weekly previews
• reminders
• scheduled spiritual material

without me remembering to ask for them.

That is where AI starts becoming quite different from a chatbot.

Instead of:

I remember something → open AI → ask AI

it becomes:

The system remembers → AI brings it to me when it matters.

---

Research and monitoring

Bud also watches various sources for me.

For example, it can:

• monitor selected YouTube channels
• monitor AI developments
• summarise new material
• filter out things that aren't particularly important
• run research jobs
• delegate longer research to specialist agents

This means AI can sometimes act less like a person I'm talking to and more like a piece of infrastructure quietly performing jobs I've previously asked it to do.

---

My FPL manager

One example is a complete Fantasy Premier League management system I've built.

There are a number of FPL creators whose analysis I find useful.

My system can collect and process their video transcripts and use that information alongside data about my own team.

That gives me a way of bringing together many hours of FPL discussion and using it to help analyse things such as:

• transfers
• captaincy
• player form
• fixtures
• injuries
• upcoming deadlines
• different creators' opinions

Rather than relying on one person's view, the system can help me see the overall picture across several different analysts.

It's essentially a specialised AI research assistant built around one particular hobby.

---

AI-assisted security camera monitoring

Another project I've built is a Reolink camera watcher for security camera monitoring during LDC work.

When you have perhaps 8–16 camera feeds visible at once, you still need to keep watching the screens properly, but anything that can draw attention to activity can be useful additional assistance.

The program allows me to select the part of the computer screen containing the camera feeds and then monitors each of those individual feeds.

It uses two layers of detection.

First, it can watch for the person-detection icon produced by the Reolink cameras themselves.

I've also integrated a YOLO computer-vision person detector, so the software can independently detect and track people appearing within the camera feeds.

So it doesn't replace the person monitoring the cameras.

It's an additional layer that can help draw attention towards one of many feeds when relevant activity is detected.

That was another project that began simply with:

"It would be useful if software could help with this."

and eventually became something I could actually build.

---

Multiple AI agents

Another area I've experimented with is multi-agent systems.

Bud doesn't necessarily have to perform every task itself.

It can delegate work to other agents specialising in things such as:

• research
• coding
• debugging
• analysis

I've experimented with tools and agents including Bud, Codex, Claude Code, OpenClaw, Antigravity and others.

That means the workflow can sometimes look like:

Me → Bud → specialist agent → result → Bud → me

instead of me manually moving everything between different AI systems.

---

About 80 specialised AI skills

I've gradually built a library of roughly 80 specialist skills.

A skill is basically a repeatable procedure teaching the AI how I want a certain type of job done.

Examples include workflows for:

• talk preparation assistance
• Bible research
• family worship ideas
• rehabilitation tracking
• fantasy football
• YouTube monitoring
• project work
• coding
• system maintenance

This is another change in how I now think about AI.

Rather than repeatedly explaining:

"Whenever I ask you to do this, please remember to do A, B, C, D and E…"

I can encode that method once as a reusable skill.

---

Around 47 scheduled jobs

The system currently operates roughly 47 scheduled jobs.

Some use AI.

Others deliberately don't.

They include things such as:

• daily briefings
• spiritual encouragement
• reminders
• monitoring
• backups
• housekeeping
• health checks
• weekly reviews
• data updates

One useful lesson has been learning when NOT to use AI.

If a normal script can reliably do something, I generally prefer the script.

AI is most useful where judgement, interpretation, language or reasoning is required.

So the architecture becomes something like:

ordinary software for predictable work + AI for intelligent work.

---

What I've actually learnt to build

Perhaps the more interesting part for me isn't Bud itself.

It's what experimenting with AI has enabled me to learn.

I wouldn't describe myself as a professional programmer or AI engineer, yet with AI helping me I have been able to design and build systems I simply would not previously have attempted.

AI system architecture

I've learnt to put together systems involving:

• a main AI assistant
• specialist sub-agents
• task delegation
• scheduled jobs
• scripts
• databases
• knowledge stores
• local servers
• APIs
• monitoring
• dashboards
• backup systems
• recovery processes

I can start with the outcome I want and use AI to help work backwards towards the architecture.

---

Knowledge-base engineering

One of the things I've enjoyed most has been creating specialist knowledge bases.

Instead of merely hoping the model knows something, I can index authoritative material and tell it:

Search this first.

That is how my spiritual research system works.

The AI can retrieve relevant material and then reason over it.

That concept can be used with almost anything:

personal notes
technical documentation
books
research papers
project histories
specialist reference material

The intelligence comes partly from the model, but also from giving the model the right information at the right time.

---

Defining an AI's behaviour

I've also learnt that building a useful AI agent isn't just about choosing the smartest model.

A huge amount depends on defining things such as:

• its role
• its tone
• its boundaries
• what sources it must consult
• what it is allowed to change
• when it must ask permission
• how it should admit uncertainty

For Bud I've maintained an identity and behaviour specification so that it stays warm and practical rather than becoming theatrical or pretending to know things it doesn't.

I've also built rules such as:

• verify scripture wording
• search the local library for doctrinal research
• don't expose secrets
• restrict external systems where appropriate
• be clear about what the agent actually knows

AI still needs boundaries.

---

AI supervising AI

I've increasingly become interested in using AI to supervise other AI.

For example:

AI architect → AI builder → AI tests → AI reviewer

or:

AI researcher → source checking → review → final synthesis

That reduces my dependence on simply trusting the first answer produced by one model.

Sometimes one of the most useful jobs for AI is checking the work produced by another AI.

---

Automation with human approval

I deliberately haven't tried to make everything fully autonomous.

For important changes I prefer:

AI proposes → I review → I approve → system implements

That seems much more sensible than simply giving an AI unrestricted permission to modify everything.

Automation is useful.

Oversight is useful too.

---

Keeping it economical

I've also put quite a bit of thought into AI cost.

Using the most powerful model for every task is unnecessary.

I'm experimenting with systems that can choose between:

• simple deterministic code
• smaller/cheaper AI models
• general models
• coding models
• stronger reasoning models

depending on what the job actually requires.

The principle is roughly:

Use the least expensive method capable of doing the job reliably.

Only escalate to the more powerful models when their additional intelligence is actually needed.

---

Other things I've built with AI

Away from Bud, I've used AI to help design and build quite a variety of things.

These include:

• the integrated fitness and rehabilitation system
• the Reolink/YOLO camera-monitoring assistant
• the complete FPL management system
• AI monitoring dashboards
• an agent observability system
• personal productivity tools
• project dashboards
• music-making software
• drum machines
• bass-line creators
• sub-bass generators
• ambient and psychedelic music tools
• workflow automation
• server monitoring
• databases
• research tools
• decision dashboards
• experimental multi-agent systems

Often my part begins with something as vague as:

"Wouldn't it be useful if something existed that could do this?"

Then I can discuss it with AI until that becomes:

idea → specification → architecture → design → code → testing → working application

That, for me, is probably one of the most remarkable parts of current AI.

The distance between having an idea and being capable of building the idea has suddenly become much smaller.

---

Research

The other major way I use AI is research.

I've used it for subjects ranging from technology, AI and software through to health, history and Bible-related research.

One particularly deep example has been my research into the Neo-Babylonian period.

I've used AI to help investigate things such as:

• Babylonian kings
• dated business tablets
• regnal chronology
• astronomical records
• historical synchronisms
• ancient historical sources
• archaeological evidence
• Scriptural passages
• academic literature

What makes AI useful here isn't simply asking it for an answer.

I can give it a large research question and then keep drilling down into particular pieces of evidence.

For example:

"What is the original source for this claim?"

"How strong is this evidence?"

"Is this source contemporary with the event?"

"Are these five websites actually relying on the same original source?"

"Can we reconstruct the sequence year by year?"

This is where AI can save an enormous amount of time while still leaving the human responsible for evaluating the evidence.

---

Source-critical research

I frequently ask AI not merely to find information, but to distinguish between different kinds of evidence.

For example:

primary evidence → contemporary records → near-contemporary records → later historical sources → modern academic interpretation → opinion

That helps enormously with subjects where twenty websites may appear to provide twenty pieces of evidence when they're actually all repeating the same underlying claim.

AI can help trace things back towards the original source.

---

Researching current information

For things that change frequently, I use AI together with live internet research.

Examples include:

• developments in AI
• new AI models
• technology news
• software releases
• products
• prices
• travel and accommodation
• services
• technical documentation

The big difference from an ordinary search engine is that I can search, compare the results, ask follow-up questions and have the material organised into something understandable.

---

Researching videos and long material

I also use AI to get useful information out of long videos and documents.

I can give it a long video or transcript and ask:

• What were the main arguments?
• What evidence was presented?
• Which parts actually matter?
• What practical ideas came out of it?
• Is there anything here that I could use or build?

Quite a few of my software projects have actually started because I've seen an idea somewhere and then asked AI:

"Could we build our own version of that?"

---

Writing and communicating clearly

One of the less spectacular but probably most useful applications of AI for me is simply writing faster and more clearly.

I use it constantly for:

• rewriting messages
• shortening long explanations
• improving structure
• removing repetition
• making something tactful
• making something warmer
• turning notes into finished prose
• organising research
• explaining complicated subjects simply
• producing summaries at different levels of detail

I can often write a rough paragraph containing exactly what I mean and ask AI to preserve the meaning while improving the communication.

That can save a surprising amount of time.

And because I do quite a lot of administrative and congregation communication, that little everyday use of AI is probably just as valuable to me as some of the much more complicated systems I've built.

---

Probably the biggest change

I don't really think of AI as a chatbot anymore.

Increasingly I think of it as a collection of capabilities:

Researcher
Programmer
Architect
Designer
Critic
Editor
Scheduler
Monitor
Knowledge librarian
Automation engine
Orchestrator

The clever part isn't necessarily any single AI model.

It's increasingly about combining:

AI + ordinary software + trusted source material + automation + databases + human oversight.

And ultimately that is what I find most useful about it.

I'm not trying to automate my life for the sake of automating it.

I'm trying to remove some of the repetitive mental load, speed up things that otherwise take unnecessary time, and make information easier to manage so that more attention can go towards the things that actually matter.

And occasionally, of course, using a ridiculously powerful collection of AI models to build something completely unnecessary simply because I thought, "That would be brilliant." 😄