diff --git a/ml/README.md b/ml/README.md
new file mode 100644
index 0000000..22fd5dd
--- /dev/null
+++ b/ml/README.md
@@ -0,0 +1,43 @@
+# Fruit maturity and quality classifiers
+
+This directory contains a PyTorch training script for the two subtasks in the dataset:
+
+- **Maturity** (`Immature`, `Mature`) stored under `Maturity/`
+- **Quality** (`Fresh`, `Rotten`) stored under `Quality/`
+
+The script handles variable input image sizes by resizing to 224×224 with torchvision transforms and saves checkpoints using the requested names `maturity_best-3-1.pth` and `quality_best-3-1.pth`.
+
+## Dataset layout
+```
+<root>
+├── Maturity/
+│   ├── Immature/
+│   └── Mature/
+└── Quality/
+    ├── Fresh/
+    └── Rotten/
+```
+
+Place your images in the matching class folders. The script automatically splits each task into train/validation sets.
+
+## Training
+Train both maturity and quality models (ResNet-18 fine-tuned with ImageNet weights by default):
+```bash
+python ml/train_fruit_models.py train --data-root /path/to/data --epochs 15 --batch-size 32 --output-dir checkpoints
+```
+
+Train only one task:
+```bash
+python ml/train_fruit_models.py train --data-root /path/to/data --task maturity
+python ml/train_fruit_models.py train --data-root /path/to/data --task quality
+```
+
+The best validation checkpoints are saved to `maturity_best-3-1.pth` and `quality_best-3-1.pth` inside `--output-dir`.
+
+## Prediction
+Run inference on a single image (device auto-selects CUDA if available during training; you can override for prediction):
+```bash
+python ml/train_fruit_models.py predict --checkpoint checkpoints/maturity_best-3-1.pth --image /path/to/image.jpg --device cpu
+```
+
+The command prints the predicted class and per-class probabilities.
diff --git a/ml/train_fruit_models.py b/ml/train_fruit_models.py
new file mode 100644
index 0000000..e9f2120
--- /dev/null
+++ b/ml/train_fruit_models.py
@@ -0,0 +1,305 @@
+"""Training and prediction utilities for fruit maturity and quality models.
+
+Dataset layout expected by this script:
+
+<root>
+├── Maturity/
+│   ├── Immature/
+│   └── Mature/
+└── Quality/
+    ├── Fresh/
+    └── Rotten/
+
+Use torchvision transforms so variable-size images are supported.
+"""
+
+from __future__ import annotations
+
+import argparse
+import random
+from collections import Counter
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Dict, Iterable, Tuple
+
+import torch
+import torch.nn as nn
+import torch.optim as optim
+from PIL import Image
+from torch.utils.data import DataLoader, random_split
+from torchvision import datasets, models, transforms
+
+
+@dataclass
+class DataConfig:
+    data_root: Path
+    output_dir: Path
+    epochs: int = 10
+    batch_size: int = 32
+    learning_rate: float = 1e-3
+    weight_decay: float = 1e-4
+    train_split: float = 0.8
+    num_workers: int = 4
+    seed: int = 42
+    use_pretrained: bool = True
+
+
+TASK_INFO = {
+    "maturity": {
+        "subdir": "Maturity",
+        "classes": ["Immature", "Mature"],
+        "checkpoint": "maturity_best-3-1.pth",
+    },
+    "quality": {
+        "subdir": "Quality",
+        "classes": ["Fresh", "Rotten"],
+        "checkpoint": "quality_best-3-1.pth",
+    },
+}
+
+
+def seed_everything(seed: int) -> None:
+    random.seed(seed)
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed_all(seed)
+
+
+def build_transforms(image_size: int = 224) -> Tuple[transforms.Compose, transforms.Compose, transforms.Compose]:
+    train_transform = transforms.Compose(
+        [
+            transforms.Resize((image_size, image_size)),
+            transforms.RandomHorizontalFlip(),
+            transforms.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2),
+            transforms.ToTensor(),
+            transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
+        ]
+    )
+
+    eval_transform = transforms.Compose(
+        [
+            transforms.Resize((image_size, image_size)),
+            transforms.ToTensor(),
+            transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
+        ]
+    )
+
+    predict_transform = eval_transform
+    return train_transform, eval_transform, predict_transform
+
+
+def split_dataset(dataset: datasets.ImageFolder, train_split: float, seed: int) -> Tuple[torch.utils.data.Dataset, torch.utils.data.Dataset]:
+    train_size = int(len(dataset) * train_split)
+    val_size = len(dataset) - train_size
+    generator = torch.Generator().manual_seed(seed)
+    return random_split(dataset, [train_size, val_size], generator=generator)
+
+
+def build_dataloaders(cfg: DataConfig, task: str) -> Tuple[DataLoader, DataLoader, Dict[int, int]]:
+    task_cfg = TASK_INFO[task]
+    train_transform, eval_transform, _ = build_transforms()
+
+    base_dataset = datasets.ImageFolder(cfg.data_root / task_cfg["subdir"], transform=train_transform)
+    train_set, val_set = split_dataset(base_dataset, cfg.train_split, cfg.seed)
+
+    val_set.dataset = datasets.ImageFolder(cfg.data_root / task_cfg["subdir"], transform=eval_transform)
+
+    class_counts = Counter([label for _, label in base_dataset])
+
+    train_loader = DataLoader(
+        train_set,
+        batch_size=cfg.batch_size,
+        shuffle=True,
+        num_workers=cfg.num_workers,
+        pin_memory=True,
+    )
+
+    val_loader = DataLoader(
+        val_set,
+        batch_size=cfg.batch_size,
+        shuffle=False,
+        num_workers=cfg.num_workers,
+        pin_memory=True,
+    )
+
+    return train_loader, val_loader, dict(class_counts)
+
+
+def build_model(num_classes: int, use_pretrained: bool = True) -> nn.Module:
+    weights = models.ResNet18_Weights.IMAGENET1K_V1 if use_pretrained else None
+    model = models.resnet18(weights=weights)
+    model.fc = nn.Linear(model.fc.in_features, num_classes)
+    return model
+
+
+def compute_accuracy(logits: torch.Tensor, targets: torch.Tensor) -> float:
+    preds = logits.argmax(dim=1)
+    correct = (preds == targets).sum().item()
+    return correct / targets.size(0)
+
+
+def class_weight_tensor(class_counts: Dict[int, int], device: torch.device) -> torch.Tensor:
+    total = sum(class_counts.values())
+    weights = [total / class_counts[i] for i in range(len(class_counts))]
+    return torch.tensor(weights, dtype=torch.float32, device=device)
+
+
+def evaluate(model: nn.Module, loader: DataLoader, criterion: nn.Module, device: torch.device) -> Tuple[float, float]:
+    model.eval()
+    total_loss = 0.0
+    total_acc = 0.0
+    with torch.no_grad():
+        for images, labels in loader:
+            images, labels = images.to(device), labels.to(device)
+            outputs = model(images)
+            loss = criterion(outputs, labels)
+            total_loss += loss.item() * images.size(0)
+            total_acc += compute_accuracy(outputs, labels) * images.size(0)
+
+    return total_acc / len(loader.dataset), total_loss / len(loader.dataset)
+
+
+def train_one_task(cfg: DataConfig, task: str, device: torch.device) -> Path:
+    seed_everything(cfg.seed)
+    task_cfg = TASK_INFO[task]
+    train_loader, val_loader, counts = build_dataloaders(cfg, task)
+
+    model = build_model(num_classes=len(task_cfg["classes"]), use_pretrained=cfg.use_pretrained)
+    model.to(device)
+
+    criterion = nn.CrossEntropyLoss(weight=class_weight_tensor(counts, device))
+    optimizer = optim.Adam(model.parameters(), lr=cfg.learning_rate, weight_decay=cfg.weight_decay)
+
+    best_acc = 0.0
+    best_path = cfg.output_dir / task_cfg["checkpoint"]
+    best_path.parent.mkdir(parents=True, exist_ok=True)
+
+    for epoch in range(1, cfg.epochs + 1):
+        model.train()
+        running_loss = 0.0
+        running_acc = 0.0
+        for images, labels in train_loader:
+            images, labels = images.to(device), labels.to(device)
+            optimizer.zero_grad()
+            outputs = model(images)
+            loss = criterion(outputs, labels)
+            loss.backward()
+            optimizer.step()
+
+            running_loss += loss.item() * images.size(0)
+            running_acc += compute_accuracy(outputs.detach(), labels) * images.size(0)
+
+        epoch_loss = running_loss / len(train_loader.dataset)
+        epoch_acc = running_acc / len(train_loader.dataset)
+
+        val_acc, val_loss = evaluate(model, val_loader, criterion, device)
+        print(
+            f"[{task}] Epoch {epoch}/{cfg.epochs} "
+            f"Train Loss: {epoch_loss:.4f} Acc: {epoch_acc:.4f} | "
+            f"Val Loss: {val_loss:.4f} Acc: {val_acc:.4f}"
+        )
+
+        if val_acc > best_acc:
+            best_acc = val_acc
+            torch.save(
+                {
+                    "model_state": model.state_dict(),
+                    "classes": task_cfg["classes"],
+                    "task": task,
+                },
+                best_path,
+            )
+            print(f"[{task}] New best model saved to {best_path} (val acc={val_acc:.4f})")
+
+    return best_path
+
+
+def predict(image_path: Path, checkpoint: Path, device: torch.device) -> Tuple[str, Dict[str, float]]:
+    if not checkpoint.exists():
+        raise FileNotFoundError(f"Checkpoint not found: {checkpoint}")
+
+    payload = torch.load(checkpoint, map_location=device)
+    classes: Iterable[str] = payload["classes"]
+    task: str = payload["task"]
+    _, _, predict_transform = build_transforms()
+
+    model = build_model(num_classes=len(classes), use_pretrained=False)
+    model.load_state_dict(payload["model_state"])
+    model.to(device)
+    model.eval()
+
+    image = Image.open(image_path).convert("RGB")
+    tensor = predict_transform(image).unsqueeze(0).to(device)
+
+    with torch.no_grad():
+        logits = model(tensor)
+        probs = torch.softmax(logits, dim=1).squeeze(0)
+
+    confidence = {cls: probs[idx].item() for idx, cls in enumerate(classes)}
+    predicted_class = classes[int(probs.argmax().item())]
+    return predicted_class, confidence
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Train fruit maturity and quality classifiers.")
+    subparsers = parser.add_subparsers(dest="command", required=True)
+
+    train_parser = subparsers.add_parser("train", help="Train models")
+    train_parser.add_argument("--data-root", type=Path, required=True, help="Dataset root containing Maturity/ and Quality/ folders")
+    train_parser.add_argument("--output-dir", type=Path, default=Path("."), help="Directory to store checkpoints")
+    train_parser.add_argument("--epochs", type=int, default=10)
+    train_parser.add_argument("--batch-size", type=int, default=32)
+    train_parser.add_argument("--learning-rate", type=float, default=1e-3)
+    train_parser.add_argument("--weight-decay", type=float, default=1e-4)
+    train_parser.add_argument("--train-split", type=float, default=0.8)
+    train_parser.add_argument("--num-workers", type=int, default=4)
+    train_parser.add_argument("--seed", type=int, default=42)
+    train_parser.add_argument("--no-pretrained", action="store_true", help="Disable ImageNet pretraining")
+    train_parser.add_argument("--task", choices=["maturity", "quality", "both"], default="both")
+
+    predict_parser = subparsers.add_parser("predict", help="Run inference on a single image")
+    predict_parser.add_argument("--checkpoint", type=Path, required=True, help="Path to a saved .pth checkpoint")
+    predict_parser.add_argument("--image", type=Path, required=True, help="Path to an RGB image")
+    predict_parser.add_argument("--device", type=str, default="cuda" if torch.cuda.is_available() else "cpu")
+
+    return parser.parse_args()
+
+
+def handle_train(args: argparse.Namespace) -> None:
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    cfg = DataConfig(
+        data_root=args.data_root,
+        output_dir=args.output_dir,
+        epochs=args.epochs,
+        batch_size=args.batch_size,
+        learning_rate=args.learning_rate,
+        weight_decay=args.weight_decay,
+        train_split=args.train_split,
+        num_workers=args.num_workers,
+        seed=args.seed,
+        use_pretrained=not args.no_pretrained,
+    )
+
+    tasks = [args.task] if args.task != "both" else ["maturity", "quality"]
+    for task in tasks:
+        print(f"Training task: {task} on device {device}")
+        best_checkpoint = train_one_task(cfg, task, device)
+        print(f"Finished training {task}. Best checkpoint: {best_checkpoint}")
+
+
+def handle_predict(args: argparse.Namespace) -> None:
+    device = torch.device(args.device)
+    label, conf = predict(args.image, args.checkpoint, device)
+    formatted = ", ".join([f"{k}: {v:.3f}" for k, v in conf.items()])
+    print(f"Prediction: {label} ({formatted})")
+
+
+def main() -> None:
+    args = parse_args()
+    if args.command == "train":
+        handle_train(args)
+    elif args.command == "predict":
+        handle_predict(args)
+
+
+if __name__ == "__main__":
+    main()