cursor @Cursor0P

We were only allowed to use gemma model for tha hackathon, so i could not able to build mobile app under 3-4GB of ram. I'm planning to use small VLM models around 1B to make it possible post hackathon. Might require some fine-tuning but will try.

cursor @Cursor0P

4 months ago

Built this for #Gemma4Hackathon on @Kaggle: AgriDoc a local-first crop disease assistant for farmers. Using Gemma 4 E2B + @ollama, it turns crop images into local-language, citation-backed guidance with what to do today / tomorrow / next 7 days. #Kaggle #Gemma #AIforGood #Agri

8 0 0 76 0

0 0 0 52 0

View Details

cursor @Cursor0P

4 months ago

I will soon share my AgriDoc projects code, once i polish it a bit. I'm hoping to build it to a level where it can be used as a real problem solving application, instead of side project for hackthaon. Due to hackathon time limit, i only build very basic MVP. #agritech #AIforGood

0 0 0 22 0

View Details

cursor @Cursor0P

4 months ago

@kaggle @ollama This way i can scale the supported disease and crops without touching the model. Only by updating the KB. Reducing the chances of LLM hallucinations, and wrong recommendations which can be fatal. SLM suffers from hallucination, grounded solutions build trust.

0 0 0 24 0

View Details

cursor @Cursor0P

4 months ago

8 0 0 76 0

View Details

cursor @Cursor0P

4 months ago

@kaggle @ollama show the top matching diseases with sample images, so the farmer could choose for himself which disease information and visualy match his crops image. Once he choose the disease, we will provide him the solution from our KB.

0 0 0 14 0

View Details

cursor @Cursor0P

4 months ago

@kaggle @ollama But due to time limit and i had only 2 days, i decided to use models vision capability to describe the image as list of symptoms. Build a KB which had crop and disease information including symptoms, match the given symptoms against the KB and search top matching disease.

0 0 0 15 0

View Details

cursor @Cursor0P

4 months ago

@kaggle @ollama I was going to use @Alibaba_Qwen vision model to prepare the dataset, as i can use them for synthetic dataset generation as they are open source models. I will be providing the images and ask it to describe the symptoms, which will be used to identify the disease.

0 0 0 10 0

View Details

cursor @Cursor0P

4 months ago

@kaggle @ollama Initial my plan was also to use @UnslothAI to fine-tune gemma4 on crop dataset, but all datasets were classifier based and importantly not with MIT license to use. I needed to build a dataset where i will have image, crop name, disease name and models reasoning how it identified.

0 0 0 9 0

View Details

cursor @Cursor0P

4 months ago

@kaggle @ollama The problem with typical CNN based disease prediction apps is, you can not scale them without retraining them. And the main problem is dataset for training on wide range of the disease all across the world for making sure it will help farmers on global level.

0 0 0 7 0

View Details

cursor @Cursor0P

4 months ago

@kaggle @ollama The main purpose of distributing app to the millions of farmers, was no longer viable. I was hoping to get this under 3-4GB of ram. So anyone with smartphone could able to use it. Gemma E2B alone was taking 4GB for quants weight.

0 0 0 9 0

View Details

cursor @Cursor0P

4 months ago

@kaggle @ollama Initially, I was building the mobile version of this using the cactus runtime. But while testing the initial MVP using Cactus-Compute/LFM2-VL-450M I faced issues of high ram usage and crashes.