You will get I will build a U-Net image segmentation model for your dataset
Project details
I build custom U-Net image segmentation models that classify
every pixel in your images — for medical imaging, satellite
analysis, drone footage, microscopy, and more.
What makes my work different:
— I built U-Net from scratch in TensorFlow achieving Mean IoU
of 0.674 across 7 land cover classes on the DeepGlobe satellite
dataset — presented at ICRTAI 2025 international conference.
— I implement Focal Tversky combined with weighted cross-entropy
loss to handle class imbalance — a critical issue most developers
ignore that destroys model performance on rare classes.
— I use Albumentations for aggressive augmentation — flip,
rotate, elastic transform — so your model generalizes well
even on small datasets.
— I deliver a working Streamlit app so you can test predictions
on new images instantly without writing any code.
What you get:
✅ Custom U-Net for your image size and number of classes
✅ Full data augmentation pipeline
✅ Custom loss function for class imbalance
✅ Per-class IoU, precision, recall evaluation
✅ Clean documented code and trained model file
✅ Streamlit demo app (Advanced tier)
Share your dataset and I will build your model.
every pixel in your images — for medical imaging, satellite
analysis, drone footage, microscopy, and more.
What makes my work different:
— I built U-Net from scratch in TensorFlow achieving Mean IoU
of 0.674 across 7 land cover classes on the DeepGlobe satellite
dataset — presented at ICRTAI 2025 international conference.
— I implement Focal Tversky combined with weighted cross-entropy
loss to handle class imbalance — a critical issue most developers
ignore that destroys model performance on rare classes.
— I use Albumentations for aggressive augmentation — flip,
rotate, elastic transform — so your model generalizes well
even on small datasets.
— I deliver a working Streamlit app so you can test predictions
on new images instantly without writing any code.
What you get:
✅ Custom U-Net for your image size and number of classes
✅ Full data augmentation pipeline
✅ Custom loss function for class imbalance
✅ Per-class IoU, precision, recall evaluation
✅ Clean documented code and trained model file
✅ Streamlit demo app (Advanced tier)
Share your dataset and I will build your model.
Machine Learning Tools
Azure Machine Learning, BERT, ChatGPT, deeplearn.js, Deeplearning4j, GitHub Copilot, GPT-3, Keras, Kubeflow, MLflow, NLTK, NumPy, Open Neural Network Exchange, OpenCV, pandas, Python, Python Scikit-Learn, PyTorch, scikit-learn, SQL, TensorFlow, XGBoostWhat's included
| Service Tiers |
Starter
$50
|
Standard
$100
|
Advanced
$150
|
|---|---|---|---|
| Delivery Time | 7 days | 8 days | 15 days |
Number of Revisions | 1 | 2 | 3 |
Number of Model Variations | 1 | 2 | 3 |
Number of Scenarios | 1 | 2 | 3 |
Number of Graphs/Charts | 2 | 4 | 6 |
Model Validation/Testing | |||
Model Documentation | - | ||
Data Source Connectivity | - | ||
Source Code |
Optional add-ons
You can add these on the next page.
Streamlit Web App
(+ 3 Days)
+$80
Custom Loss Function
(+ 2 Days)
+$60Frequently asked questions
About Yamraj
Machine Learning & AI Enthusiast
Dharan, Nepal - 5:38 am local time
I build production ML and AI systems — not just notebooks.
I'm a Computer Engineering student at IOE Purwanchal Campus (Nepal) specializing in Machine Learning, LLMs, Computer Vision, and RAG pipelines. My work has been adopted by the open-source community and presented at international AI conferences.
WHAT I'VE BUILT
— Fine-tuned Mistral-7B on Nepal's National Penal Code → adopted by community maintainer within 24 hours and redistributed in Q2–Q8 GGUF formats for CPU-only inference across the llama.cpp ecosystem
— Built U-Net from scratch in TensorFlow for satellite land cover segmentation (Mean IoU: 0.674 across 7 classes) → presented at ICRTAI 2025 international conference
— Built multi-agent AI system using LangChain + LangGraph with 5 specialized agents: Mood Detection, FAISS-based Memory (RAG), Content Generation, Surprise Planning, Safety Check
— Deployed 4 live AI web apps + 2 production FastAPI endpoints + Android app (React Native/Expo) on Hugging Face Spaces
— Implemented RNN, GRU, LSTM, and BRNN architectures from scratch in PyTorch
— Built MLOps pipeline: MLflow experiment tracking, Optuna hyperparameter search, model registry
WHAT I CAN DO FOR YOU
1.Fine-tune LLMs on your domain-specific data
2. Build RAG pipelines for your documents or knowledge base
3.Train computer vision models (classification, segmentation, detection)
4. Build and deploy ML APIs with FastAPI on Hugging Face or cloud
5. MLOps: experiment tracking, reproducible pipelines, model registry
6. Multi-agent AI workflows with LangChain and LangGraph
TECH STACK
Python | TensorFlow | PyTorch | Hugging Face | FAISS | FastAPI | Streamlit | LangChain | LangGraph | MLflow | GGUF / llama.cpp | React Native
Steps for completing your project
After purchasing the project, send requirements so Yamraj can start the project.
Delivery time starts when Yamraj receives requirements from you.
Yamraj works on your project following the steps below.
Revisions may occur after the delivery date.
Analyze dataset and plan architecture
I review your images, masks and class distribution. I check for imbalance, assess image quality, and confirm the best U-Net configuration for your task within 24 hours.
Build preprocessing and augmentation pipeline
I build resizing, normalization and augmentation pipeline using Albumentations — horizontal flip, vertical flip, rotation, elastic transform and brightness contrast adjustments.