How local AI improved your live?
Posted by Thin_Pollution8843@reddit | LocalLLaMA | View on Reddit | 38 comments
Lets share use cases which improve life quality of the people. Home assistants, psychological help, local coding, deep reasearch, business help etc.
I personally working rn on a local health tracker. PDFs with bloodwork in - structurised data out which I will use later to analyse and track separate blood params. Still thinking about how to incorporate Docs conclusions/ultrasound/ECGs results or images etc in to that.
grabber4321@reddit
Its basically like a brainy junior developer that knows a lot of shit, but when it puts it together you have to fix every step so its coherent.
sometimes_angery@reddit
Bold for you to assume it did
comp21@reddit
I run an AI for medical research (this is aside from my genetic research).
The AI has my genome file, my supplements and medications spreadsheet with dosage, what I take, what I used to take, time of day/week I take them and when I started them.
Back in Jan I noticed I had been having minor tremors in my pinky fingers and eye lids. I kept track of symptoms: started in Oct, starts when I wake up. Intensity doesn't change but duration is... What ended at 930am before now stops around 2pm... Put it all in AI and it spit back:
"Based on your SCOLB1 gene (I might have this wrong, this is by memory) you process Crestor improperly which is causing your high dosage to build up in your bloodstream and causing damage to your nerves as it's absorbed by your muscles. Your timeline of when you started matches well with when statin toxicity symptoms begin. You should ask your doctor to about switching to pravastatin. You do also have a gene that changes how you process pravastatin however research is split on whether the gene makes it more or less effective"
I told my doc my symptoms and the gene I have. He agreed and... Switched me to pravastatin (I did not tell him about the AI or the pravastatin suggestion). I've been on it for 90 days now and feel WAY better. Had it continued I'd have ended up with irreversible nerve damage.
Thin_Pollution8843@reddit (OP)
Very cool! I also have my genome sequenced could you tell a bit about your setup? How you make model to use your genome and other stuff?
comp21@reddit
To add to my original response: most of my work has been in finding ways to cut out noise, false positives etc. Depending on who did your genome there could be all sorts of errors (for example, my 30x from sequencing.com shows me having a genetic defect that would have, 100%, killed me in infancy... Since I am older than three months old I'm going to assume either I don't have that or it's not "on").
That's the hard part...and don't get me started on 23andme files...
comp21@reddit
Oh dude this has been a 4 month project. My scripts are now, all combined, over 60 pages long and I pull data from five public genetic databases.
I would be happy to run you a report if you like. Just pm me. I plan to make this in to a website where people can run their own at some point. I think I'm getting down to where the reports are useful but I'm running ones for whoever wants them right now to get feedback on how to improve them.
arbv@reddit
That is genuinely one of the best uses for AI I have seen. What models did you use?
comp21@reddit
I originally used Gemini pro under my Google one subscription but that was before the nerfing. Now I'm transitioning over to a local medgemma:27b-q8 model. I have been using my Gemini sub to run questions on both models to test medgemma 's accuracy and so far it's been right on point.
arbv@reddit
I hope they will produce Gemma 4-based version.
SaucyParamecium@reddit
How do you actually communicate this to your doctor? I found that the medical community is highly resilient to exchanging and listening patients . If I ever go to any doctor I know and say "I have read this on a paper since I'm having those symptoms and I would like to investigate it bla bla" I'd get insta shut with him replying that "Dr Google is never right ". I think most of them have some complex.
With that said, I would find it nearly impossible to ask him to find an AI discovered
comp21@reddit
I did not in any way tell him AI told me this but I do have a very very good cardiologist who is extremely research oriented. I just message him through the hospital message system.
It might help that he did misdiagnose me before. It was me who asked for a stress test which showed my afib was not pacemaker related but actual blockages (and my genetic research started then... Which has made me find three genes that caused my early calcification... We're working on new treatments now that I have this info). He does tend to listen to me a lot however he's also the type who will so there's that.
Long story short: my triple bypass happened a few years ago. The cardio I'm with now moved to another hospital two hours away so I took his recommendation for a local doc. The local doc was like you're describing: a dick. So I switched back to my old doc and just drive two hours.
Technical-Earth-3254@reddit
It's fun
Celestialien@reddit
Slightly nerdier answer, but I write LLM pipelines and evals were always the bottleneck for me. Scoring the outputs with a frontier model as the judge adds up surprisingly fast when you're running it over thousands of rows every single time you change a prompt, so I ended up doing it way less than I should have just to keep the bill down.
Swapping the judge out for a local model basically killed that. I hand it the output along with a rubric and it gives me back a score and a short reason, and while it's obviously not as sharp as a big model, it's more than good enough to flag when a change has made things worse, which is most of what I actually need it for anyway. The bigger thing is that since it costs nothing to run I'll happily grade the whole set three or four times in an afternoon while I'm iterating, whereas before I'd be sat there rationing API calls and telling myself a spot check of ten outputs was probably fine.
mystery_biscotti@reddit
Dude, that's awesome. 😎
mkMoSs@reddit
Qwen3.6 27B turned out to be a fantastic coding assistant. I'm working from home and I feel like I literally have hired an intern to do a lot of the coding grunt work. Always under my supervision, no code "leaves" without my review.
I run it on VLLM. I use it with VSCodium and Roo (Zoo) Code extension. I also use the codebase indexing feature of Roo (Zoo) Code which the LLM can use to easily search for things in my codebase.
Hardware: 4x RTX 5060 Ti 16gb so 64GB vram.
havnar-@reddit
It makes for a google messed up search.
Can’t say it’s been an improvement. I just need it to compete in my field or get left behind.
Guilty-Guitar-9366@reddit
Local LLM tagged and categorized over 1,000 image files for me. If I didn’t have this, I’d probably still be working on this task. I was also able to extract valuable information about old content that I hadn’t even bothered to look for because it was buried under the worst kind of hate speech. Even if it runs a little slow, there’s nothing better than having it do this automatically while I drink my coffee and eat a cookie.
JLeonsarmiento@reddit
My room is hotter during the winter.
Mean-Ad1493@reddit
Just having fun and marveling at the capacity of my potato PC. Nothing productive so far though
Former_Bathroom_2329@reddit
Now I'm spending all weekend on llm homeland.
jopereira@reddit
Without these new tools, I wouldn't start some protects as they would be too complex to be worth it. Now, I can do ESP32 projects for chips I've bought in 2018 and never used them... With web UI, projects are way more complex and competent but still easy to implement as a simple Arduino Nano project (much less flexible, less powerful, less... everything!!).
StrongZeroSinger@reddit
I can ask it for the 300th time how to restart and rebuild a docker container without having to get misleading commands from google or wasting tokens on enterprise models lol
UnethicalExperiments@reddit
I finally have my multivac at home. While still getting the hang of the tools and agents, I have a sys admin in a box. I've been waiting for this since i was 8 wondering why my parents were so pissed about the commodore in pieces while because I "wanted to see what made it think" after reading Asimov for the first time .
I keep telling people, we are just scratching the surface here of what this can be. The more of us using it at home , the faster this will accelerate.
DonkeyBonked@reddit
Local AI has allowed me to learn to build, pre-train, fine-tune, and customize models with LoRAs, MCP servers, Tools, Skills, RAG/Vector databases, and building my own agent framework. It is what gives me skills in AI to build AI systems that work how I want them to and make money doing it for others.
After a lifetime in tech, local open-source AI is what allowed me to transition from being a developer who was trying to keep up and learn to adapt my workflow to use AI, to an AI developer.
I see far more potential in local AI than cloud crap, local is the future we all need.
Hot-Employ-3399@reddit
It can code in background while I watch YouTube. I sometimes review changes, tell what to change and can go back to not bother too much over something that is necessary. (Sometimes I also tell agents to review changes)
keepthepace@reddit
I must confess that for coding I'm still using remotely hosted agents, but I have taken the habit for two years to mostly speak to them and I rarely type more than a paragraph of text, thanks to whisperX.
Just like this comment that I don't care to type directly.
qudat@reddit
I’ve been trying to internalize and articulate how an agent is truly useful without me being lazy. I think there’s a version of agent use that is lazy: I want this built but I don’t want to do it. Then there’s another version of agent use where I’m blocked (I don’t know how to solve a problem) or writers blocked (stuck on a part I don’t want to continue). Those have been amazing for me. There’s a bunch of sysadmin stuff I just don’t enjoy, for example, having the agent convert podman quadlet containers to docker. That’s something I wanted to do but it would have taken me forever to build up the courage to do it solo.
Thebandroid@reddit
I only need to invest another $1200 and I’ll have save $240 in AI subscriptions this year!
soyalemujica@reddit
I've burned about 10m tokens in Qwen 3.6 27B Q5KM in pure backend and frontend programming for my game server including 2.5D map editor, it has been perfectly functional, no need to pay for a frontier with my needs
finah1995@reddit
It has allowed increased access to code and migrate older stuff. And also as local AI it's more easier to help not worrying about sharing know-how and intellectual property to third-party service providers. Also health wise and planning some things.
I use third-party AI but either for general things like questions, designing systems or increasing my knowledge. Or even generating conceptual system architecture and breaking them and evaluating them.
I use the similar structures and knowledge from there with local LLM and make do in works but I don't give organizational knowledge over to the third parties.
Sadale-@reddit
It's just fun to tinker with it. I'm not using it for anything productive for now. :D
Public_Parfait_6412@reddit
Also this!
BigYoSpeck@reddit
Being able to experiment, build and iterate apps that use them locally without worrying about API costs is how I learnt the skills needed to land my current job
Public_Parfait_6412@reddit
This!
Public_Parfait_6412@reddit
I like building workflow pipelines that one off and specific model for specific tasks. Like n8n but with no fees. And once you have off grid or web ui, can hit it from your phone.
autoencoder@reddit
I use it as a diary and sometimes coding assistant. Kinda sorta 24/7 psychological help, but I'm aware it's just a realistic parrot, and I use it to cheer for me to do my "homework" from a real psychologist.
Also works for navigating social situations. Sometimes there's tension or drama, and it lets me understand more of other people's possible motives.
Same as you, I wouldn't share any personal stuff with Big Tech. Friends included.
asertym@reddit
I've fixed and launched 4 personal projects that I started as a hobby years ago and never got to finish, I'm more organized since I'm taking more notes and it's easier than ever to go through them, I've learned better English (I'm from Europe so not my native language) by writing more than before and learned to express myself better. My work related input is more structured and am spending less time on trivial tasks than before. I don't feel like I work less, but I am more productive.
szansky@reddit
Data analysis is amazing :)