diff --git a/elasticai/creator/ir/__init__.py b/elasticai/creator/ir/__init__.py
new file mode 100644
index 00000000..6473ebae
--- /dev/null
+++ b/elasticai/creator/ir/__init__.py
@@ -0,0 +1,6 @@
+from .abstract_ir_data import AbstractIRData, MandatoryField
+
+__all__ = [
+    "MandatoryField",
+    "AbstractIRData",
+]
diff --git a/elasticai/creator/ir/abstract_ir_data/__init__.py b/elasticai/creator/ir/abstract_ir_data/__init__.py
new file mode 100644
index 00000000..d5c375de
--- /dev/null
+++ b/elasticai/creator/ir/abstract_ir_data/__init__.py
@@ -0,0 +1,7 @@
+from .abstract_ir_data import AbstractIRData
+from .mandatory_field import MandatoryField
+
+__all__ = [
+    "AbstractIRData",
+    "MandatoryField",
+]
diff --git a/elasticai/creator/ir/abstract_ir_data/_attributes_descriptor.py b/elasticai/creator/ir/abstract_ir_data/_attributes_descriptor.py
new file mode 100644
index 00000000..0383b79a
--- /dev/null
+++ b/elasticai/creator/ir/abstract_ir_data/_attributes_descriptor.py
@@ -0,0 +1,16 @@
+from collections.abc import MutableMapping
+
+from elasticai.creator.ir.attribute import AttributeT
+
+from ._has_data import HasData
+from ._hiding_dict import _HidingDict
+
+
+class _AttributesDescriptor:
+    def __init__(self, hidden_names: set[str]):
+        self._hidden_names = hidden_names
+
+    def __get__(
+        self, instance: HasData, owner: type[HasData]
+    ) -> MutableMapping[str, AttributeT]:
+        return _HidingDict(self._hidden_names, instance.data)
diff --git a/elasticai/creator/ir/abstract_ir_data/_has_data.py b/elasticai/creator/ir/abstract_ir_data/_has_data.py
new file mode 100644
index 00000000..b5ff2652
--- /dev/null
+++ b/elasticai/creator/ir/abstract_ir_data/_has_data.py
@@ -0,0 +1,9 @@
+from typing import Protocol, runtime_checkable
+
+from elasticai.creator.ir.attribute import AttributeT
+
+
+@runtime_checkable
+class HasData(Protocol):
+    @property
+    def data(self) -> dict[str, AttributeT]: ...
diff --git a/elasticai/creator/ir/abstract_ir_data/_hiding_dict.py b/elasticai/creator/ir/abstract_ir_data/_hiding_dict.py
new file mode 100644
index 00000000..59684fac
--- /dev/null
+++ b/elasticai/creator/ir/abstract_ir_data/_hiding_dict.py
@@ -0,0 +1,83 @@
+import collections
+from collections.abc import Iterable
+from itertools import filterfalse
+from typing import MutableMapping, TypeVar
+
+T = TypeVar("T")
+
+
+class _HidingDict(MutableMapping[str, T]):
+    """Allows to hide keys with `hidden_names` for all read operations.
+    We use this to implement an attributes field for Nodes that looks like a dictionary, but hides
+    all mandatory fields.
+    You can still write to `HidingDict`, e.g.,
+
+    >>> d = dict(a="a", b="b")
+    >>> h = _HidingDict({"a"}, d)
+    >>> "b", == tuple(h.keys())
+    True
+    >>> "a" in h
+    False
+    >>> h["a"] = "c"
+    >>> h.data["a"]
+    'c'
+    >>> d["a"]
+    'c'
+    >>> "a" in d and "a" in h.data
+    True
+    """
+
+    def __init__(self, hidden_names: Iterable[str], data: dict) -> None:
+        self.data = data
+        self._hidden_names = set(hidden_names)
+
+    def __setitem__(self, key: str, value: T):
+        self.data[key] = value
+
+    def _is_hidden(self, name: str) -> bool:
+        return name in self._hidden_names
+
+    def __delitem__(self, key: str):
+        del self.data[key]
+
+    def __iter__(self):
+        return filterfalse(self._is_hidden, iter(self.data))
+
+    def __contains__(self, item):
+        # overriding this should also make class behave correctly for getting items
+        return item not in self._hidden_names and item in self.data
+
+    def __getitem__(self, item: str) -> T:
+        return self.data[item]
+
+    def __len__(self):
+        return len(self.data)
+
+    def get(self, key: str, default=None) -> T:
+        if key in self:
+            return self[key]
+        return default
+
+    def __copy__(self) -> MutableMapping[str, T]:
+        inst = self.__class__.__new__(self.__class__)
+        inst.__dict__.update(self.__dict__)
+        # Create a copy and avoid triggering descriptors
+        inst.__dict__["data"] = self.__dict__["data"].copy()
+        return inst
+
+    def copy(self) -> MutableMapping[str, T]:
+        if self.__class__ is collections.UserDict:
+            return _HidingDict(self._hidden_names.copy(), self.data.copy())
+        import copy
+
+        data = self.data
+        try:
+            self.data = {}
+            c = copy.copy(self)
+        finally:
+            self.data = data
+        c.update(self)
+        return c
+
+    def __repr__(self) -> str:
+        return f"HidingDict({', '.join(self._hidden_names)}, data={self.data})"
diff --git a/elasticai/creator/ir/abstract_ir_data/abstract_ir_data.py b/elasticai/creator/ir/abstract_ir_data/abstract_ir_data.py
new file mode 100644
index 00000000..6a9c7859
--- /dev/null
+++ b/elasticai/creator/ir/abstract_ir_data/abstract_ir_data.py
@@ -0,0 +1,171 @@
+import inspect
+import sys
+from abc import abstractmethod
+from collections.abc import Iterable
+from typing import Any, Callable, TypeVar
+
+from elasticai.creator.ir.attribute import AttributeT
+
+from ._attributes_descriptor import _AttributesDescriptor
+from .mandatory_field import MandatoryField, TransformableMandatoryField
+
+if sys.version_info.minor > 10:
+    from typing import Self
+else:
+    Self = TypeVar("Self", bound="AbstractIrData")
+
+
+class AbstractIRData:
+    """
+    This class should provide a way to easily create new wrappers around dictionaries.
+
+    It is supposed to be used together with the `MandatoryField` class.
+    Every child of `AbstractIRData` is expected to have a constructor that takes a dictionary.
+    That dictionary is not copied, but instead can be shared with other Node classes.
+
+    Most of the private functions in this class deal with handling arguments of the classmethod
+    `new`, that is used to create new nodes from scratch.
+
+    The `attributes` attribute of the class provides a dict-like object, that hides all keys that are associated
+    with mandatory fields.
+
+    The purpose of this class is to provide a way to easily write new wrappers around dictionaries, that let us customize
+    access, while still allowing static type annotations.
+    """
+
+    __slots__ = ("data",)
+    attributes = _AttributesDescriptor(set())
+
+    def __init__(self: Self, data: dict[str, AttributeT]):
+        """IMPORTANT: Do not override this. If you want to create a function that creates new nodes of your subtype,
+        override the `new` method instead.
+        """
+        for k in self._mandatory_fields():
+            if k not in data:
+                raise ValueError(f"Missing mandatory field {k}")
+        self.data = data
+
+    @classmethod
+    def _do_new(cls, *args, **kwargs) -> Self:
+        """This is here for your convenience to be called in `new`."""
+        cls.__validate_arguments(args, kwargs)
+        data = cls.__turn_arguments_into_data_dict(args, kwargs)
+        return cls(data)
+
+    @classmethod
+    @abstractmethod
+    def new(cls, *args, **kwargs) -> Self:
+        """Create a new node by creating a new dictionary from args and kwargs.
+
+        Use this for creation of new nodes from inline code. This is typically also where you want to provide
+        type hints for users via `@overload`. You can delegate to the `_do_new()` method of `BaseNode`
+        """
+        ...
+
+    def as_dict(self: Self) -> dict[str, AttributeT]:
+        return self.data
+
+    @classmethod
+    def from_dict(cls: type[Self], data: dict[str, AttributeT]) -> Self:
+        return cls(data)
+
+    def __eq__(self: Self, other: object) -> bool:
+        if hasattr(other, "data") and isinstance(other.data, dict):
+            return self.data == other.data
+        else:
+            return False
+
+    @classmethod
+    def __turn_arguments_into_data_dict(
+        cls, args: tuple[Any], kwargs: dict[str, Any]
+    ) -> dict[str, Any]:
+        data = cls.__extract_attributes_from_args(args, kwargs)
+        data.update(cls.__get_kwargs_without_attributes(kwargs))
+        data.update(cls.__get_args_as_kwargs(args))
+        cls.__transform_args_with_mandatory_fields(data)
+        return data
+
+    @classmethod
+    def __get_mandatory_field_descriptors(
+        cls,
+    ) -> Iterable[tuple[str, TransformableMandatoryField]]:
+        for c in reversed(inspect.getmro(cls)):
+            for a in c.__dict__:
+                if (
+                    not a.startswith("__")
+                    and not a.endswith("__")
+                    and isinstance(c.__dict__[a], TransformableMandatoryField)
+                ):
+                    yield a, c.__dict__[a]
+
+    @classmethod
+    def _mandatory_fields(cls) -> tuple[str, ...]:
+        return tuple(name for name, _ in cls.__get_mandatory_field_descriptors())
+
+    def __attribute_keys(self: Self):
+        return tuple(k for k in self.data.keys() if k not in self._mandatory_fields())
+
+    def __repr__(self: Self):
+        mandatory_fields_repr = ", ".join(
+            f"{k}={self.data[k]}" for k in self._mandatory_fields()
+        )
+        attributes = ", ".join(
+            f"'{k}': '{self.data[k]}'" for k in self.__attribute_keys()
+        )
+        return (
+            f"{self.__class__.__name__}({mandatory_fields_repr},"
+            f" attributes={attributes})"
+        )
+
+    @classmethod
+    def __validate_arguments(cls, args: tuple[Any], kwargs: dict[str, Any]):
+        num_total_args = len(args) + len(kwargs)
+        if num_total_args not in (
+            len(cls._mandatory_fields()),
+            len(cls._mandatory_fields()) + 1,
+        ):
+            raise ValueError(
+                f"allowed arguments are {cls._mandatory_fields()} and attributes, but"
+                f" passed args: {args} and kwargs: {kwargs}"
+            )
+        argument_names_in_args = set(k for k, _ in zip(cls._mandatory_fields(), args))
+        arguments_specified_twice = argument_names_in_args.intersection(kwargs.keys())
+        if len(arguments_specified_twice) > 0:
+            raise ValueError(f"arguments specified twice {arguments_specified_twice}")
+
+    @classmethod
+    def __extract_attributes_from_args(
+        cls, args: tuple[Any, ...], kwargs: dict[str, Any]
+    ) -> dict[str, Any]:
+        data = dict()
+        if "attributes" in kwargs:
+            if callable(kwargs["attributes"]):
+                data.update(kwargs["attributes"]())
+            else:
+                data.update(kwargs["attributes"])
+        elif len(args) + len(kwargs) == len(cls._mandatory_fields()) + 1:
+            attributes = args[-1]
+            data.update(attributes)
+        return data
+
+    @classmethod
+    def __get_kwargs_without_attributes(
+        cls, kwargs: dict[str, Any]
+    ) -> dict[str, AttributeT]:
+        kwargs = {k: v for k, v in kwargs.items() if k != "attributes"}
+        return kwargs
+
+    @classmethod
+    def __transform_args_with_mandatory_fields(cls, args: dict[str, Any]) -> None:
+        set_transforms = cls.__get_field_transforms()
+        for k in args:
+            if k in set_transforms:
+                args[k] = set_transforms[k](args[k])
+
+    @classmethod
+    def __get_field_transforms(cls) -> dict[str, Callable[[Any], AttributeT]]:
+        return {k: v.set_transform for k, v in cls.__get_mandatory_field_descriptors()}
+
+    @classmethod
+    def __get_args_as_kwargs(cls, args: tuple[Any]) -> Iterable[tuple[str, Any]]:
+        return zip(cls._mandatory_fields(), args)
diff --git a/elasticai/creator/ir/abstract_ir_data/mandatory_field.py b/elasticai/creator/ir/abstract_ir_data/mandatory_field.py
new file mode 100644
index 00000000..93bab39d
--- /dev/null
+++ b/elasticai/creator/ir/abstract_ir_data/mandatory_field.py
@@ -0,0 +1,72 @@
+from collections.abc import Callable
+from typing import cast
+
+from typing_extensions import Generic, Protocol, TypeVar
+
+from elasticai.creator.ir.attribute import AttributeT
+
+from ._has_data import HasData
+
+T = TypeVar("T", bound=AttributeT)  # stored data type
+
+F = TypeVar("F", default=T)  # visible data type
+
+
+class AbstractIR(Protocol):
+    data: dict[str, AttributeT]
+
+
+class TransformableMandatoryField(Generic[T, F]):
+    """
+    A __descriptor__ that designates a mandatory field of an abstract ir data class.
+    The descriptor accesses the `data` dictionary of the owning abstract ir data object
+    to read and write values. You can use the `set_transform` and `get_transform` functions
+    to transform values during read/write accesses. `T` designates the type stored in the
+    `data` dictionary, while `F` is the type that the mandatory field receives.
+    That allows to keep dictionary of primitive (serializable) data types in memory,
+    while still providing abstract ways to manipulate that data in complex ways.
+    This is typically required when working with Nodes and Graphs to create new
+    intermediate representations and transform one graph into another.
+
+    E.g.
+
+    ```python
+    class A(AbstractIrData):
+        number: TransformableMandatoryField[str, int] = TransformableMandatoryField(set_transform=str, get_transform=int)
+
+
+    a = A({'number': "12"})
+    a.number = a.number + 3
+    print(a.data) # {'number': "15"}
+    ```
+    """
+
+    def __init__(
+        self,
+        set_transform: Callable[[F], T],
+        get_transform: Callable[[T], F],
+    ):
+        self.set_transform = set_transform
+        self.get_transform = get_transform
+
+    def __set_name__(self, owner, name: str) -> None:
+        """
+        IMPORTANT: do not remove owner even though it's not used
+        see https://docs.python.org/3/reference/datamodel.html#descriptors for more information
+        """
+        self.name = name
+
+    def __get__(self, instance: HasData, owner) -> F:
+        """
+        IMPORTANT: do not remove owner even though it's not used
+        see https://docs.python.org/3/reference/datamodel.html#descriptors for more information
+        """
+        return self.get_transform(cast(T, instance.data[self.name]))
+
+    def __set__(self, instance: HasData, value: F) -> None:
+        instance.data[self.name] = self.set_transform(value)
+
+
+class MandatoryField(TransformableMandatoryField[T, T]):
+    def __init__(self):
+        super().__init__(lambda x: x, lambda x: x)
diff --git a/elasticai/creator/ir/attribute.py b/elasticai/creator/ir/attribute.py
new file mode 100644
index 00000000..2441d7e3
--- /dev/null
+++ b/elasticai/creator/ir/attribute.py
@@ -0,0 +1,6 @@
+from typing import TypeAlias
+
+SizeT: TypeAlias = tuple[int] | tuple[int, int] | tuple[int, int, int]
+AttributeT: TypeAlias = (
+    int | float | str | tuple["AttributeT", ...] | dict[str, "AttributeT"]
+)
diff --git a/elasticai/creator/ir/node.py b/elasticai/creator/ir/node.py
new file mode 100644
index 00000000..c550f2d5
--- /dev/null
+++ b/elasticai/creator/ir/node.py
@@ -0,0 +1,25 @@
+from typing import overload
+
+from .abstract_ir_data import AbstractIRData
+from .abstract_ir_data.mandatory_field import MandatoryField
+from .attribute import AttributeT
+
+
+class Node(AbstractIRData):
+    name: MandatoryField[str] = MandatoryField()
+    type: MandatoryField[str] = MandatoryField()
+
+    @classmethod
+    @overload
+    def new(cls, name: str, type: str, attributes: dict[str, AttributeT]) -> "Node": ...
+
+    @classmethod
+    @overload
+    def new(cls, name: str, type: str) -> "Node": ...
+
+    @classmethod
+    def new(cls, *args, **kwargs) -> "Node":
+        return cls._do_new(*args, **kwargs)
+
+    def __hash__(self):
+        return hash((self.name, self.type))
diff --git a/elasticai/creator/ir/node_test.py b/elasticai/creator/ir/node_test.py
new file mode 100644
index 00000000..f9d472af
--- /dev/null
+++ b/elasticai/creator/ir/node_test.py
@@ -0,0 +1,114 @@
+from typing import overload
+
+import pytest
+
+from elasticai.creator.ir.abstract_ir_data.mandatory_field import MandatoryField
+
+from .abstract_ir_data import AbstractIRData
+from .abstract_ir_data.mandatory_field import TransformableMandatoryField
+from .node import Node
+from .node import Node as OtherNode
+
+
+def test_can_create_new_node():
+    n = Node.new(name="my_name", type="my_type")
+    assert n.name == "my_name"
+
+
+def test_creating_new_node_without_filling_all_args_yields_error():
+    with pytest.raises(ValueError):
+        Node.new(name="my_name")
+
+
+def test_filling_name_arg_twice_leads_to_error():
+    with pytest.raises(ValueError):
+        Node.new("my_name", "my_type", name="other_name")
+
+
+def test_attributes_are_merged_into_data():
+    n = Node.new(name="my_name", type="my_type", attributes=dict(input_shape=(1, 1)))
+    assert n.data == dict(name="my_name", type="my_type", input_shape=(1, 1))
+
+
+def test_attributes_are_merged_into_data_for_call_with_positional_args_only():
+    n = Node.new("my_name", "my_type", dict(input_shape=(1, 1)))
+    assert n.data == dict(name="my_name", type="my_type", input_shape=(1, 1))
+
+
+def test_two_different_node_types_can_share_data() -> None:
+    class NewNode(AbstractIRData):
+        input_shape: MandatoryField[tuple[int, int]] = MandatoryField()
+
+        @classmethod
+        @overload
+        def new(cls, input_shape: tuple[int, int]) -> "NewNode": ...
+
+        @classmethod
+        def new(cls, *args, **kwargs) -> "NewNode":
+            return cls._do_new(*args, **kwargs)
+
+    a = Node.new(name="my_name", type="my_type", attributes=dict(input_shape=(1, 3)))
+    b = NewNode(a.data)
+    b.input_shape = (4, 3)
+    assert a.attributes["input_shape"] == b.input_shape
+
+
+def test_nodes_inherit_mandatory_fields() -> None:
+    class NewNode(Node):
+        input_shape: MandatoryField[tuple[int, int]] = MandatoryField()
+
+        @classmethod
+        @overload
+        def new(
+            cls, name: str, type: str, input_shape: tuple[int, int]
+        ) -> "NewNode": ...
+
+        @classmethod
+        def new(cls, *args, **kwargs) -> "NewNode":
+            return cls._do_new(*args, **kwargs)
+
+    n = NewNode.new(name="my_name", type="my_type", input_shape=(1, 3))
+    assert n.input_shape == (1, 3)
+
+
+def test_can_serialize_node_with_attributes():
+    n = Node.new(name="my_node", type="my_type", attributes={"a": "b", "c": (1, 2)})
+    assert dict(name="my_node", type="my_type", a="b", c=(1, 2)) == n.as_dict()
+
+
+def test_can_deserialize_node_with_attributes():
+    n = Node.new(name="my_node", type="my_type", attributes={"a": "b", "c": (1, 2)})
+    assert n == Node.from_dict(dict(name="my_node", type="my_type", a="b", c=(1, 2)))
+
+
+def test_import_path_does_not_matter_for_equality():
+    n = Node.new(name="a", type="a_type", attributes=dict(a="b"))
+    other = OtherNode.new(name="a", type="a_type", attributes=dict(a="b"))
+    assert n == other
+
+
+def test_set_transform_is_applied_when_calling_new() -> None:
+    def set_transform(x: str) -> int:
+        return int(x)
+
+    def get_transform(x: int) -> str:
+        return str(x)
+
+    class NewNode(Node):
+        input_shape: TransformableMandatoryField[int, str] = (
+            TransformableMandatoryField(
+                set_transform=set_transform,
+                get_transform=get_transform,
+            )
+        )
+
+        @classmethod
+        @overload
+        def new(cls, name: str, type: str, input_shape: str) -> "NewNode": ...
+
+        @classmethod
+        def new(cls, *args, **kwargs) -> "NewNode":
+            return cls._do_new(*args, **kwargs)
+
+    n = NewNode.new(name="my_name", type="my_type", input_shape="3")
+    assert n.input_shape == "3"
diff --git a/elasticai/creator/ir/typing.py b/elasticai/creator/ir/typing.py
new file mode 100644
index 00000000..6062af87
--- /dev/null
+++ b/elasticai/creator/ir/typing.py
@@ -0,0 +1,22 @@
+from typing import Protocol, Self, TypeVar, runtime_checkable
+
+from .attribute import AttributeT
+
+
+@runtime_checkable
+class Node(Protocol):
+    name: str
+    type: str
+    data: dict[str, AttributeT]
+    attributes: dict[str, AttributeT]
+
+    @classmethod
+    def new(cls, *args, **kwargs) -> Self: ...
+
+    @classmethod
+    def from_dict(cls, d: dict[str, AttributeT]) -> Self: ...
+
+    def as_dict(self) -> dict[str, AttributeT]: ...
+
+
+NodeT = TypeVar("NodeT", bound=Node)