diff --git a/v2-roadmap.md b/v2-roadmap.md
new file mode 100644
index 00000000..554f88ff
--- /dev/null
+++ b/v2-roadmap.md
@@ -0,0 +1,201 @@
+# V2 implementation proposal
+
+## Why this proposal
+
+The current implementation uses different `CloudEvent` classes for the implementation
+of different protocol bindings. This adds complexity the integration of new technologies
+like `pydantic`.
+
+The separation of responsibilities in the package is not clear. We have multiple JSON
+implementations, while the format specification is the same for all bindings in the
+[`CloudEvent` spec](https://github.com/cloudevents/spec/blob/v1.0.2/cloudevents/formats/json-format.md):
+
+* Pydantic implementation is calling the HTTP implementation for this, adding unnecessary overhead
+* the Kafka binding uses its own implementation, creating code duplication and unnecessary maintenance efforts
+
+We should have a single `CloudEvent` class used for all bindings/formats since the event
+is the Python representation of the event data, which is the same no matter what binding
+is used. (we potentially could receive a HTTP event and forward it as it is using Kafka).
+
+The specific bindings and formats implementations should be separate from the `CloudEvent`
+class.
+
+This proposal contains some base rules for the package architecture and how we could implement
+the `CloudEvent` spec in this package.
+
+## Requirements
+
+In order to be able to transform a python CloudEvent to/from a specific
+protocol binding, this SDK has the following responsibilities:
+
+* Identifying what attributes have to be marshalled/unmarshalled (identifying the cloudevents extensions)
+* Validate required fields and generation of defaults
+* Marshalling/unmarshalling the cloudevents core and extensions fields
+* Marshalling/unmarshalling the `data` field
+* Encode/decode the marshalled fields to/from the specific binding format
+
+## Modules
+
+We should separate the implementation of the requirements so that they can be tested independently and
+each component be used for composition.
+
+Ideally we should have these modules (naming TBC):
+
+* `cloudevents.event` - This would contain the `CloudEvent` base class. This class is the same for all bindings/formats.
+* `cloudevents.formats` - This would contain the structured formats implementations (JSON, AVRO, Protobuf)
+* `cloudevents.bindings` - This would contain the specific binding implementations (kafka, http)
+
+This matches the naming and concepts we have in the `CloudEvent` spec
+
+## Matching requirements and modules
+
+### The `cloudevents.event` module and `CloudEvent` class
+
+The `CloudEvent` class would satisfy the following requirements:
+
+* Validate required fields and generation of default values
+* Marshalling/unmarshalling the `data` field
+* Identifying what attributes have to be marshalled/unmarshalled
+
+The `data` field depends on the specific event and the `contentdatatype` attribute and
+its marshalling/unmarshalling logic should be the same for any binding/format. It seems
+a `CloudEvent` class responsibility to take care of this. (We should implement this logic
+in overridable methods)
+
+Ideally the end user will be able to extend the `CloudEvent` class for their necessity. Ie:
+
+* adding new extensions fields
+* adding fields/methods not meant to be marshalled (internal fields)
+* implementing new defaults (ie. creating a `MyEvent` class with `type="myevent"` default)
+
+We also want a DTO that can be accepted by the most common libraries (ie. JSON) to keep the
+`CloudEvent` class decoupled from the formats/bindings. Using a dictionary seems a good idea.
+
+We should implement a `to_dict`/`from_dict` functionalities in the `CloudEvent` class.
+
+QUESTION: Can we define a data type for the `data` field after it's been marshalled?
+(ie. can we say it will be either `str` or binary data?)
+
+### The `cloudevents.formats` module
+
+Formats do solve a single responsibility:
+
+* Marshalling/unmarshalling the cloudevents core and extensions fields
+
+Their implementation has to be used when other bindings implement their
+structured input/output and make sure their implementation is compatible.
+
+Each format accepts the dictionary from the `CloudEvent` as param
+and produces an output based on the format itself. Ie.
+
+```python
+def to_json(event: dict) -> str: ...
+def from_json(event: str) -> dict: ...
+```
+
+TODO: Review typing as JSON produces a string but other formats might be different.
+
+### The `cloudevents.bindings` module
+
+Bindings do solve these responsibilities:
+
+* Marshalling/unmarshalling the cloudevents core and extensions fields **(if binary format)**
+* Encode/decode the marshalled fields to/from the specific binding format
+
+They also have the responsibility of coordinating the other modules.
+
+This is a **concept** example for `cloudevents.bindings.http` module
+
+```python
+from cloudevents.formats.json import to_json, from_json
+from cloudevents.event import CloudEvent
+
+
+# Shared dataclass to help with typing, might be a dictionary or a named tuple (not important)
+@dataclass
+class HTTPCloudEvent:
+    headers: list
+    body: Any  # unsure about typing here
+
+
+def to_binary(event: CloudEvent) -> HTTPCloudEvent:
+    e = event.to_dict()
+    # logic to marshall into http binary output
+    return HTTPCloudEvent(headers=[], body=[])
+
+
+def to_json(event: CloudEvent) -> HTTPCloudEvent:
+    e = to_json(event.to_dict())
+    # logic to compose the JSON response
+    return HTTPCloudEvent(headers=[], body=[])
+
+
+def from_binary(headers: list, body: dict) -> CloudEvent:
+    e = {}
+    # specific logic to unmarshall raw headers into a dictionary
+    return CloudEvent.from_dict(e)
+
+
+def from_json(body: str) -> CloudEvent:
+    e = from_json(body)
+    return CloudEvent.from_dict(e)
+
+
+def from_http(headers: list, body: dict) -> CloudEvent:
+    # identify if structured or binary based on headers / body
+    # and use the proper `from_` function
+    # We should also handle here batches
+    ...
+```
+
+Other bindings would behave similarly, where the end user would need only
+to use the relevant module in `cloudevents.bindings` namespace.
+
+## CloudEvent subclasses
+
+We'll want SDK users to be able to extend the `CloudEvent` class, so they will
+be able to manage marshalling/unmarshalling of the `data` field for different
+event types and their python typing.
+
+This could be easily done by introducing a mapping between the `type` field and
+`CloudEvent` subclasses as a parameter in the `from_` functions. Ie.
+
+```python
+def from_binary(headers: list, body: dict, event_types: Optional[dict[str, type]] = None) -> CloudEvent:
+    event_as_dict = {}
+    # specific logic to unmarshall raw headers into a dictionary
+    if event_types:
+        # Use
+        return event_types.get(event_as_dict['type'], CloudEvent).from_dict(e)
+    else:
+        return CloudEvent.from_dict(e)
+```
+
+## Pydantic support
+
+The current Pydantic implementation is implemented as a possible substitute of
+HTTP Event class, by looking at its conversion module and tests, but it's used
+by calling the functions in conversion module.
+
+Pydantic offers good support and performance for data validation, defaults and
+JSON serialization (some of the requirements we have identified)
+
+We need to make a decision:
+
+* We use pydantic as a base for `CloudEvent` class:
+  * PROs
+    * No need to implement data validation
+    * Performance (pydantic V2 is written in Rust)
+    * We can lay down some base logic on `data` field serialization (ie. nested pydantic models)
+    * Enable users to use pydantic for automatic documentation functionalities
+  * CONs
+    * We introduce a new package dependency, and depend on its implementation
+    * _minor_ If we use pydantic JSON serialization it is implemented in the model class (slightly different from what proposed but we can use wrapper functions in the JSON format module)
+* We remove pydantic support:
+  * PROs
+    * We have more control over validation and defaults implementation
+    * We don't depend on other packages
+  * CONs
+    * Performance on validation and defaults (we can still use libraries like `orjson` for JSON performance)
+    * Additional effort in writing and maintaining data validation and required/default fields
+...