@azure

Devlog 11: i was right llms were gonna have an issue: tokens. but so far i;ve completely fixed Machine Learning which was always saying it was deforestation due to it being a u-net classifier. which was pretty pointless so i added before and after images and used mask to tell if deforestation occurred and accounting for if trees grew back it wouldn’t just kill system, i added cot functional, and llm was overthinking but i;ve been working on that lately. im in the home stretch for shipping all i want to do is expand where it scans, and chunking the scans so it doesn’t use too many resources, and possibly upgrading to a database instead of json? ill do that in ship #2 as it’ll add more bands and more. but im gonna widen scanning. at mini um. also just identified a bottle neck with my permit tester so ill fix that as well its messing with llm.

Devlog 10.75 had to fix left over placeholder values, gave up on Sentile-1 for this current ship because couldn’t find datasets and i don’t have time or compute to make my own. and I have been working on frontend, and fixing some task. and added a scan button in frontend for interactivity. im now testing the scans and making sure it actually sends over images as i’ve finally debugged ALMOST everything. i had a hiccup with model because i forgot the scope so it got angry there and i have a feeling llm will give me problems so ill need to test and see if it hallucinates.

Devlog 10.5: adding new machine learning, about to train, and added SAR satellite and going to train wont finish till tmrw or 2 days most likely.

Devlog 10: finally go it mainly functional. while it doesn’t look that way im gonna fix that, its just a json not fully adding + API issue on react side and i removed map when removing dummy data so i have to add it back. but other than that fully functional as i know. Some issues i ran into while debugging were, i stupidly forgot to link my .env and google earth engine key to docker, and bigger, i didn’t have a concurrency argument inside of celery workers so i ran into OOM form 96 workers. besides that just some name mix matches, i normalized the time across everything. and just debugged ngnix server to rout correctly, i left it cached and didn’t realize it was broken till i recreated to image. I plan to add sentile-1 SAR satellite to this and cleaning up frontend before i finish the ship. I decied that alerts/messing will come in ship 2 along with a more refined model and possibly more features, contacting NGOs or authorizes is a nice to have but i don’t know how’d i do it? i don’t wanna but a domain for an email and many don’t have apis, slackbot could work if they have any but that will be for ship #2 unless i change my mind.

Devlog 1: Set up the basics and made a rip off logo
nothing eventful, i set up bubble tui, nodes(cats) and pods(paws) types and go, making the protobuffer for hearbeats and stuff next

Devlog 9 : after 10 years of debugging i realized i had to make a sub domain(maybe i should’ve read docs). but after i fixed the nginix server the app was functional i just needed a subdomain to point at it. besides that i had a view errors like apikey instead of api_key, and directors being missing, and a big one requirements.txt being to big, what i did was use a swap file so its not all in memory and build service by service. another issue still on going is celery snit working which is half of my architecture and how i need to test. i think its because i set up config wrong + wrong urls/ ill try to fix it by 4am today because ill miss tmrw and possibly Friday as well(rip to my streak)

Devlog 8: Docker, so i spent a while debugging, fixing some pipelines with machine learning and sending images and more(static files) over to frontend before i added an api instead of fake data, but now the 2GB of ram isn’t looking like its enough. Before that though i had some issues setting up nginx so ill check nest documents to see if I’m doing anything wrong before i request more resources. ill try to fix instillation or find another way

Devlog 7.5: revamped frontend to show Where, and before and after, no point of plugging in api yet since i need to deploy before i can even run backend. also added call llm, and agentic loop so that should now be functional. no hiccups or issues just passing time while i wait for nest approval. And since like i said i have yet to run backend once since my computer cant handle it im expecting 1-2 hours more of debugging after i deploy.

Devlog 7: i fixxed some memory issues like a leak and just adding a garbage collector and clearing cache, also fixed ML call confidence was incorrect i think so i changed to a sigmoid, i also added EVI to differentiate vegetation and added a forest mask so NVDI should ignore if grasslands vanishes or similar. also added RGB bands separately so i can show the before and after on frontend as other bands wouldnt help + some scaling so they SHOULD appear correctly. added a health endpoint for ram and cpu usage.

Some Issues: i didn’t really need to optimize memory i didn’t think nest would have resources like it doe so i was trying to optimize for free back ends with limited ram, downtime, and such, but nest has all of that but still good to not waste memory. Second more important issue is that Brazil has a lot of clouds, so does the whole world but I’m starting in Brazil, and that will crash/ ruin data because cloud mask will just return a bunch of 0s when there are clouds, so currently i decided to do a median composite over a month, but that’s not good for real time so ill add sentile-1 to go through clouds, but ill need to research how to pair them together well and how sentile-1 can differentiate vegetation or know if they are cut down. And I have yet to plug in api to frontend so still filler data:

devlog 6: I’M A GENIUS(not really but it worked). we finally inference-d with model after training its pretty good, though in the scans I’m gonna pre-process so that it doesn’t confuse grass and trees but its not too bad at that. also added unit test though not complete, renamed a few files, and added a new endpoint to get around downtime/hibernating backed. it took a bit too long for small devlog but next ill finish unit test, dockerfile, likely retrain(possibly with more bands), and optimize so i don’t get any OOM with PyTorch, model and google earth engine. but no hiccups besides i don’t really know what to do so i just did random stuff like unit test, and also visualizing so i can verify my self was a minor hiccup. I’m gonna deploy on backend and so i can run model and earth engine and see how resource intensive it is. maybe implement sentile-1 as well to get around clouds, as you can see the cloud leaves a void with no predictions and illegal loggers use that to advantage in Brazil and log in rainy season.

Devlog 5.5: Ai is currently training so made frontend with some mock data. I’m thinking about condensing it and using a map, maybe further into it. and ofc plugging the actual complete API into it. going to do some other bug fixes and possibly testing. no real issues besides with tailwind, i had Gemini make this frontend so i can plug in API later and apparently tailwind V4 is very angry so i had to downgrade and some other setup stuff in CSS/configs.

Devlog 5: cleaned up ugly architecture and opted for celery task queuing and just calling functions and feeding it to AI, better SOC(separation of concerns) i deleted and refactored various files. such as main.py it now can scan on call while i have a scheduler daily to scan and find deforestation aync. i found a dataset for ML that may work as its GEOtiff, on HF, and has around 500 entries for training? not the best but should work as prototype, ill check for others later. some issues i had was the former agent.py was just circular and dependencies in code base were all over the place so i had to standardize it. another one was that Machine learning wasn’t working out because as mentioned previously i was scammed with kaggle dataset i found or its focused on sentile-1 instead of 2/satellite images. next devlog i plan to have a prototype/toy model, and test frontend pushing to it, and likely more debugging or finishing up/adding features, for now i’m going to the gym.

Devlog 4: I SUCK at coding. i made a task queue for machine learning but we have a couple of issues, One my pipeline for ML/Agents is just all over the place i really need to clean that up not end of world, 2. i didn’t finish push to dashboard or LLM_client. 3. the 11GB kaggle dataset SUCKKSSS. i was wondering why batch sizes were so small it has like no useful data so i need to pivot where i get data, hopefully i can find a better dataset but if worse comes to worse ill use GEE iteratively. good progress was on the queue and slightly cleaning up messy pipeline but i now need to find new data, likely change how dataset is formatted, retrain model, and finish agents and finally make pipeline make sense, for now ill work on pipeline while looking for data.

Devlog 3: Added the training loop in the notebook, fixed a few spelling errors, and added activator function(ReLU). added functionality to agent.py besides call_llm and push to dashboard. and added routing for agents. ill work on training(and likely debugging the model) and finish up agents and routing for frontend.

Made the draft of notebook/model. had chat GPT make a dashboard, and cleaned up other parts. planning to create training loop and likely debug next, and then add agent abilities.

Setting up base architecture, cant begin training till data(11 GB -_-) downloads, and im too lazy to make model before hand till i see what data looks like, gonna use kaggle api next time

Final finished the playlist issue now it should run fine, music works, slight ui changes nothing too noticeable.

removed refrences to backend opting for invisible i frame as i couldnt figure it out, it now has quing and autoplay/loops videos now. working on cleaning up ui, and ill try to get titles of music.