diff --git a/CMakeLists.txt b/CMakeLists.txt
new file mode 100644
index 000000000..bda3778bb
--- /dev/null
+++ b/CMakeLists.txt
@@ -0,0 +1,40 @@
+cmake_minimum_required(VERSION 2.8)
+
+project(numbuf)
+
+list(APPEND CMAKE_MODULE_PATH ${PROJECT_SOURCE_DIR}/cmake/Modules)
+
+find_package(PythonInterp REQUIRED)
+find_package(PythonLibs REQUIRED)
+find_package(NumPy REQUIRED)
+
+include_directories("${PYTHON_INCLUDE_DIRS}")
+include_directories("${NUMPY_INCLUDE_DIR}")
+
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11")
+
+if (UNIX AND NOT APPLE)
+   link_libraries(rt)
+endif()
+
+set(ARROW_DIR "${CMAKE_SOURCE_DIR}/../arrow/" CACHE STRING
+  "Path of the arrow source directory")
+
+set(ARROW_STATIC_LIB "${CMAKE_SOURCE_DIR}/../arrow/cpp/build/debug/libarrow.so" CACHE STRING
+  "Path to libarrow.a (needs to be changed if arrow is build in debug mode)")
+
+include_directories("${ARROW_DIR}/cpp/src/")
+include_directories("cpp/src/")
+include_directories("python/src/")
+
+add_definitions(-fPIC)
+
+add_library(numbuf SHARED
+  cpp/src/numbuf/tensor.cc
+  cpp/src/numbuf/dict.cc
+  cpp/src/numbuf/sequence.cc
+  python/src/pynumbuf/numbuf.cc
+  python/src/pynumbuf/adapters/numpy.cc
+  python/src/pynumbuf/adapters/python.cc)
+
+target_link_libraries(numbuf ${ARROW_STATIC_LIB} ${PYTHON_LIBRARIES})
diff --git a/cmake/Modules/FindNumPy.cmake b/cmake/Modules/FindNumPy.cmake
new file mode 100644
index 000000000..6e1f3c415
--- /dev/null
+++ b/cmake/Modules/FindNumPy.cmake
@@ -0,0 +1,54 @@
+# - Find the NumPy libraries
+# This module finds if NumPy is installed, and sets the following variables
+# indicating where it is.
+#
+#
+#  NUMPY_FOUND               - was NumPy found
+#  NUMPY_VERSION             - the version of NumPy found as a string
+#  NUMPY_VERSION_MAJOR       - the major version number of NumPy
+#  NUMPY_VERSION_MINOR       - the minor version number of NumPy
+#  NUMPY_VERSION_PATCH       - the patch version number of NumPy
+#  NUMPY_VERSION_DECIMAL     - e.g. version 1.6.1 is 10601
+#  NUMPY_INCLUDE_DIR         - path to the NumPy include files
+
+unset(NUMPY_VERSION)
+unset(NUMPY_INCLUDE_DIR)
+
+if(PYTHONINTERP_FOUND)
+  execute_process(COMMAND "${PYTHON_EXECUTABLE}" "-c"
+    "import numpy as n; print(n.__version__); print(n.get_include());"
+    RESULT_VARIABLE __result
+    OUTPUT_VARIABLE __output
+    OUTPUT_STRIP_TRAILING_WHITESPACE)
+
+  if(__result MATCHES 0)
+    string(REGEX REPLACE ";" "\\\\;" __values ${__output})
+    string(REGEX REPLACE "\r?\n" ";"    __values ${__values})
+    list(GET __values 0 NUMPY_VERSION)
+    list(GET __values 1 NUMPY_INCLUDE_DIR)
+
+    string(REGEX MATCH "^([0-9])+\\.([0-9])+\\.([0-9])+" __ver_check "${NUMPY_VERSION}")
+    if(NOT "${__ver_check}" STREQUAL "")
+      set(NUMPY_VERSION_MAJOR ${CMAKE_MATCH_1})
+      set(NUMPY_VERSION_MINOR ${CMAKE_MATCH_2})
+      set(NUMPY_VERSION_PATCH ${CMAKE_MATCH_3})
+      math(EXPR NUMPY_VERSION_DECIMAL
+        "(${NUMPY_VERSION_MAJOR} * 10000) + (${NUMPY_VERSION_MINOR} * 100) + ${NUMPY_VERSION_PATCH}")
+      string(REGEX REPLACE "\\\\" "/"  NUMPY_INCLUDE_DIR ${NUMPY_INCLUDE_DIR})
+    else()
+     unset(NUMPY_VERSION)
+     unset(NUMPY_INCLUDE_DIR)
+     message(STATUS "Requested NumPy version and include path, but got instead:\n${__output}\n")
+    endif()
+  endif()
+else()
+  message(STATUS "To find NumPy Python interpretator is required to be found.")
+endif()
+
+include(FindPackageHandleStandardArgs)
+find_package_handle_standard_args(NumPy REQUIRED_VARS NUMPY_INCLUDE_DIR NUMPY_VERSION
+                                        VERSION_VAR   NUMPY_VERSION)
+
+if(NUMPY_FOUND)
+  message(STATUS "NumPy ver. ${NUMPY_VERSION} found (include: ${NUMPY_INCLUDE_DIR})")
+endif()
diff --git a/cpp/src/numbuf/dict.cc b/cpp/src/numbuf/dict.cc
new file mode 100644
index 000000000..ff5f7239f
--- /dev/null
+++ b/cpp/src/numbuf/dict.cc
@@ -0,0 +1,23 @@
+#include "dict.h"
+
+using namespace arrow;
+
+namespace numbuf {
+
+std::shared_ptr<arrow::StructArray> DictBuilder::Finish(
+    std::shared_ptr<Array> list_data,
+    std::shared_ptr<Array> tuple_data,
+    std::shared_ptr<Array> dict_data) {
+  // lists and dicts can't be keys of dicts in Python, that is why for
+  // the keys we do not need to collect sublists
+  auto keys = keys_.Finish(nullptr, nullptr, nullptr);
+  auto vals = vals_.Finish(list_data, tuple_data, dict_data);
+  auto keys_field = std::make_shared<Field>("keys", keys->type());
+  auto vals_field = std::make_shared<Field>("vals", vals->type());
+  auto type = std::make_shared<StructType>(std::vector<FieldPtr>({keys_field, vals_field}));
+  std::vector<ArrayPtr> field_arrays({keys, vals});
+  DCHECK(keys->length() == vals->length());
+  return std::make_shared<StructArray>(type, keys->length(), field_arrays);
+}
+
+}
diff --git a/cpp/src/numbuf/dict.h b/cpp/src/numbuf/dict.h
new file mode 100644
index 000000000..5c2abeacd
--- /dev/null
+++ b/cpp/src/numbuf/dict.h
@@ -0,0 +1,48 @@
+#ifndef NUMBUF_DICT_H
+#define NUMBUF_DICT_H
+
+#include <arrow/api.h>
+
+#include "sequence.h"
+
+namespace numbuf {
+
+/*! Constructing dictionaries of key/value pairs. Sequences of
+    keys and values are built separately using a pair of
+    SequenceBuilders. The resulting Arrow representation
+    can be obtained via the Finish method.
+*/
+class DictBuilder {
+public:
+  DictBuilder(arrow::MemoryPool* pool = nullptr)
+    : keys_(pool), vals_(pool) {}
+
+  //! Builder for the keys of the dictionary
+  SequenceBuilder& keys() { return keys_; }
+  //! Builder for the values of the dictionary
+  SequenceBuilder& vals() { return vals_; }
+
+  /*! Construct an Arrow StructArray representing the dictionary.
+      Contains a field "keys" for the keys and "vals" for the values.
+
+      \param list_data
+        List containing the data from nested lists in the value
+        list of the dictionary
+
+      \param dict_data
+        List containing the data from nested dictionaries in the
+        value list of the dictionary
+  */
+  std::shared_ptr<arrow::StructArray> Finish(
+    std::shared_ptr<arrow::Array> list_data,
+    std::shared_ptr<arrow::Array> tuple_data,
+    std::shared_ptr<arrow::Array> dict_data);
+
+private:
+  SequenceBuilder keys_;
+  SequenceBuilder vals_;
+};
+
+}
+
+#endif
diff --git a/cpp/src/numbuf/sequence.cc b/cpp/src/numbuf/sequence.cc
new file mode 100644
index 000000000..ae250180f
--- /dev/null
+++ b/cpp/src/numbuf/sequence.cc
@@ -0,0 +1,167 @@
+#include "sequence.h"
+
+using namespace arrow;
+
+namespace numbuf {
+
+SequenceBuilder::SequenceBuilder(MemoryPool* pool)
+    : pool_(pool), types_(pool), offsets_(pool),
+      nones_(pool, std::make_shared<NullType>()),
+      bools_(pool, std::make_shared<BooleanType>()),
+      ints_(pool), strings_(pool, std::make_shared<StringType>()),
+      floats_(pool), doubles_(pool),
+      uint8_tensors_(std::make_shared<UInt8Type>(), pool),
+      int8_tensors_(std::make_shared<Int8Type>(), pool),
+      uint16_tensors_(std::make_shared<UInt16Type>(), pool),
+      int16_tensors_(std::make_shared<Int16Type>(), pool),
+      uint32_tensors_(std::make_shared<UInt32Type>(), pool),
+      int32_tensors_(std::make_shared<Int32Type>(), pool),
+      uint64_tensors_(std::make_shared<UInt64Type>(), pool),
+      int64_tensors_(std::make_shared<Int64Type>(), pool),
+      float_tensors_(std::make_shared<FloatType>(), pool),
+      double_tensors_(std::make_shared<DoubleType>(), pool),
+      list_offsets_({0}), tuple_offsets_({0}), dict_offsets_({0}) {}
+
+#define UPDATE(OFFSET, TAG)                    \
+  if (TAG == -1) {                             \
+    TAG = num_tags;                            \
+    num_tags += 1;                             \
+  }                                            \
+  RETURN_NOT_OK(offsets_.Append(OFFSET));      \
+  RETURN_NOT_OK(types_.Append(TAG));           \
+  RETURN_NOT_OK(nones_.AppendToBitmap(true));
+
+Status SequenceBuilder::Append() {
+  RETURN_NOT_OK(offsets_.Append(0));
+  RETURN_NOT_OK(types_.Append(0));
+  return nones_.AppendToBitmap(false);
+}
+
+Status SequenceBuilder::Append(bool data) {
+  UPDATE(bools_.length(), bool_tag);
+  return bools_.Append(data);
+}
+
+Status SequenceBuilder::Append(int64_t data) {
+  UPDATE(ints_.length(), int_tag);
+  return ints_.Append(data);
+}
+
+Status SequenceBuilder::Append(uint64_t data) {
+  UPDATE(ints_.length(), int_tag);
+  return ints_.Append(data);
+}
+
+Status SequenceBuilder::Append(const char* data, int32_t length) {
+  UPDATE(strings_.length(), string_tag);
+  return strings_.Append(data, length);
+}
+
+Status SequenceBuilder::Append(float data) {
+  UPDATE(floats_.length(), float_tag);
+  return floats_.Append(data);
+}
+
+Status SequenceBuilder::Append(double data) {
+  UPDATE(doubles_.length(), double_tag);
+  return doubles_.Append(data);
+}
+
+#define DEF_TENSOR_APPEND(NAME, TYPE, TAG)                                       \
+  Status SequenceBuilder::Append(const std::vector<int64_t>& dims, TYPE* data) { \
+    UPDATE(NAME.length(), TAG);                                                  \
+    return NAME.Append(dims, data);                                              \
+  }
+
+DEF_TENSOR_APPEND(uint8_tensors_, uint8_t, uint8_tensor_tag);
+DEF_TENSOR_APPEND(int8_tensors_, int8_t, int8_tensor_tag);
+DEF_TENSOR_APPEND(uint16_tensors_, uint16_t, uint16_tensor_tag);
+DEF_TENSOR_APPEND(int16_tensors_, int16_t, int16_tensor_tag);
+DEF_TENSOR_APPEND(uint32_tensors_, uint32_t, uint32_tensor_tag);
+DEF_TENSOR_APPEND(int32_tensors_, int32_t, int32_tensor_tag);
+DEF_TENSOR_APPEND(uint64_tensors_, uint64_t, uint64_tensor_tag);
+DEF_TENSOR_APPEND(int64_tensors_, int64_t, int64_tensor_tag);
+DEF_TENSOR_APPEND(float_tensors_, float, float_tensor_tag);
+DEF_TENSOR_APPEND(double_tensors_, double, double_tensor_tag);
+
+Status SequenceBuilder::AppendList(int32_t size) {
+  UPDATE(list_offsets_.size() - 1, list_tag);
+  list_offsets_.push_back(list_offsets_.back() + size);
+  return Status::OK();
+}
+
+Status SequenceBuilder::AppendTuple(int32_t size) {
+  UPDATE(tuple_offsets_.size() - 1, tuple_tag);
+  tuple_offsets_.push_back(tuple_offsets_.back() + size);
+  return Status::OK();
+}
+
+Status SequenceBuilder::AppendDict(int32_t size) {
+  UPDATE(dict_offsets_.size() - 1, dict_tag);
+  dict_offsets_.push_back(dict_offsets_.back() + size);
+  return Status::OK();
+}
+
+#define ADD_ELEMENT(VARNAME, TAG)                 \
+  if (TAG != -1) {                                \
+    types[TAG] = VARNAME.type();                  \
+    children[TAG] = VARNAME.Finish();             \
+    ARROW_CHECK_OK(nones_.AppendToBitmap(true));  \
+  }
+
+#define ADD_SUBSEQUENCE(DATA, OFFSETS, BUILDER, TAG, NAME)                    \
+  if (DATA) {                                                                 \
+    DCHECK(DATA->length() == OFFSETS.back());                                 \
+    auto list_builder = std::make_shared<ListBuilder>(pool_, DATA);           \
+    auto field = std::make_shared<Field>(NAME, list_builder->type());         \
+    auto type = std::make_shared<StructType>(std::vector<FieldPtr>({field})); \
+    auto lists = std::vector<std::shared_ptr<ArrayBuilder>>({list_builder});  \
+    StructBuilder builder(pool_, type, lists);                                \
+    OFFSETS.pop_back();                                                       \
+    ARROW_CHECK_OK(list_builder->Append(OFFSETS.data(), OFFSETS.size()));     \
+    builder.Append();                                                         \
+    ADD_ELEMENT(builder, TAG);                                                \
+  } else {                                                                    \
+    DCHECK(OFFSETS.size() == 1);                                              \
+  }
+
+std::shared_ptr<DenseUnionArray> SequenceBuilder::Finish(
+  std::shared_ptr<Array> list_data,
+  std::shared_ptr<Array> tuple_data,
+  std::shared_ptr<Array> dict_data) {
+
+  std::vector<TypePtr> types(num_tags);
+  std::vector<ArrayPtr> children(num_tags);
+
+  ADD_ELEMENT(bools_, bool_tag);
+  ADD_ELEMENT(ints_, int_tag);
+  ADD_ELEMENT(strings_, string_tag);
+  ADD_ELEMENT(floats_, float_tag);
+  ADD_ELEMENT(doubles_, double_tag);
+
+  ADD_ELEMENT(uint8_tensors_, uint8_tensor_tag);
+
+  ADD_ELEMENT(int8_tensors_, int8_tensor_tag);
+  ADD_ELEMENT(uint16_tensors_, uint16_tensor_tag);
+  ADD_ELEMENT(int16_tensors_, int16_tensor_tag);
+  ADD_ELEMENT(uint32_tensors_, uint32_tensor_tag);
+
+  ADD_ELEMENT(int32_tensors_, int32_tensor_tag);
+  ADD_ELEMENT(uint64_tensors_, uint64_tensor_tag);
+  ADD_ELEMENT(int64_tensors_, int64_tensor_tag);
+
+  ADD_ELEMENT(float_tensors_, float_tensor_tag);
+  ADD_ELEMENT(double_tensors_, double_tensor_tag);
+
+  ADD_SUBSEQUENCE(list_data, list_offsets_, list_builder, list_tag, "list");
+  ADD_SUBSEQUENCE(tuple_data, tuple_offsets_, tuple_builder, tuple_tag, "tuple");
+  ADD_SUBSEQUENCE(dict_data, dict_offsets_, dict_builder, dict_tag, "dict");
+
+  TypePtr type = TypePtr(new DenseUnionType(types));
+
+  return std::make_shared<DenseUnionArray>(type, types_.length(),
+           children, types_.data(), offsets_.data(),
+           nones_.null_count(), nones_.null_bitmap());
+}
+
+}
diff --git a/cpp/src/numbuf/sequence.h b/cpp/src/numbuf/sequence.h
new file mode 100644
index 000000000..f30ea00fe
--- /dev/null
+++ b/cpp/src/numbuf/sequence.h
@@ -0,0 +1,138 @@
+#ifndef NUMBUF_LIST_H
+#define NUMBUF_LIST_H
+
+#include <arrow/api.h>
+#include <arrow/types/union.h>
+#include "tensor.h"
+
+namespace numbuf {
+
+/*! A Sequence is a heterogeneous collections of elements. It can contain
+    scalar Python types, lists, tuples, dictionaries and tensors.
+*/
+class SequenceBuilder {
+ public:
+  SequenceBuilder(arrow::MemoryPool* pool = nullptr);
+
+  //! Appending a none to the sequence
+  arrow::Status Append();
+
+  //! Appending a boolean to the sequence
+  arrow::Status Append(bool data);
+
+  //! Appending an int64_t to the sequence
+  arrow::Status Append(int64_t data);
+
+  //! Appending an uint64_t to the sequence
+  arrow::Status Append(uint64_t data);
+
+  //! Appending a string to the sequence
+  arrow::Status Append(const char* data, int32_t length);
+
+  //! Appending a float to the sequence
+  arrow::Status Append(float data);
+
+  //! Appending a double to the sequence
+  arrow::Status Append(double data);
+
+  /*! Appending a tensor to the sequence
+
+      \param dims
+        A vector of dimensions
+
+      \param data
+        A pointer to the start of the data block. The length of the data block
+        will be the product of the dimensions
+  */
+  arrow::Status Append(const std::vector<int64_t>& dims, uint8_t* data);
+  arrow::Status Append(const std::vector<int64_t>& dims, int8_t* data);
+  arrow::Status Append(const std::vector<int64_t>& dims, uint16_t* data);
+  arrow::Status Append(const std::vector<int64_t>& dims, int16_t* data);
+  arrow::Status Append(const std::vector<int64_t>& dims, uint32_t* data);
+  arrow::Status Append(const std::vector<int64_t>& dims, int32_t* data);
+  arrow::Status Append(const std::vector<int64_t>& dims, uint64_t* data);
+  arrow::Status Append(const std::vector<int64_t>& dims, int64_t* data);
+  arrow::Status Append(const std::vector<int64_t>& dims, float* data);
+  arrow::Status Append(const std::vector<int64_t>& dims, double* data);
+
+  /*! Add a sublist to the sequenc. The data contained in the sublist will be
+     specified in the "Finish" method.
+
+     To construct l = [[11, 22], 33, [44, 55]] you would for example run
+     list = ListBuilder();
+     list.AppendList(2);
+     list.Append(33);
+     list.AppendList(2);
+     list.Finish([11, 22, 44, 55]);
+     list.Finish();
+
+     \param size
+       The size of the sublist
+  */
+  arrow::Status AppendList(int32_t size);
+
+  arrow::Status AppendTuple(int32_t size);
+
+  arrow::Status AppendDict(int32_t size);
+
+  //! Finish building the sequence and return the result
+  std::shared_ptr<arrow::DenseUnionArray> Finish(
+    std::shared_ptr<arrow::Array> list_data,
+    std::shared_ptr<arrow::Array> tuple_data,
+    std::shared_ptr<arrow::Array> dict_data);
+
+ private:
+  arrow::MemoryPool* pool_;
+
+  arrow::Int8Builder types_;
+  arrow::Int32Builder offsets_;
+
+  arrow::NullArrayBuilder nones_;
+  arrow::BooleanBuilder bools_;
+  arrow::Int64Builder ints_;
+  arrow::StringBuilder strings_;
+  arrow::FloatBuilder floats_;
+  arrow::DoubleBuilder doubles_;
+
+  UInt8TensorBuilder uint8_tensors_;
+  Int8TensorBuilder int8_tensors_;
+  UInt16TensorBuilder uint16_tensors_;
+  Int16TensorBuilder int16_tensors_;
+  UInt32TensorBuilder uint32_tensors_;
+  Int32TensorBuilder int32_tensors_;
+  UInt64TensorBuilder uint64_tensors_;
+  Int64TensorBuilder int64_tensors_;
+  FloatTensorBuilder float_tensors_;
+  DoubleTensorBuilder double_tensors_;
+
+  std::vector<int32_t> list_offsets_;
+  std::vector<int32_t> tuple_offsets_;
+  std::vector<int32_t> dict_offsets_;
+
+  int8_t bool_tag = -1;
+  int8_t int_tag = -1;
+  int8_t string_tag = -1;
+  int8_t float_tag = -1;
+  int8_t double_tag = -1;
+
+  int8_t uint8_tensor_tag = -1;
+  int8_t int8_tensor_tag = -1;
+  int8_t uint16_tensor_tag = -1;
+  int8_t int16_tensor_tag = -1;
+  int8_t uint32_tensor_tag = -1;
+  int8_t int32_tensor_tag = -1;
+  int8_t uint64_tensor_tag = -1;
+  int8_t int64_tensor_tag = -1;
+  int8_t float_tensor_tag = -1;
+  int8_t double_tensor_tag = -1;
+
+  int8_t list_tag = -1;
+  int8_t tuple_tag = -1;
+  int8_t dict_tag = -1;
+
+  int8_t num_tags = 0;
+};
+
+} // namespace numbuf
+
+#endif // NUMBUF_LIST_H
diff --git a/cpp/src/numbuf/tensor.cc b/cpp/src/numbuf/tensor.cc
new file mode 100644
index 000000000..f0f7647de
--- /dev/null
+++ b/cpp/src/numbuf/tensor.cc
@@ -0,0 +1,50 @@
+#include "tensor.h"
+
+using namespace arrow;
+
+namespace numbuf {
+
+template<typename T>
+TensorBuilder<T>::TensorBuilder(const TypePtr& dtype, MemoryPool* pool)
+    : dtype_(dtype) {
+  dim_data_ = std::make_shared<Int64Builder>(pool);
+  dims_ = std::make_shared<ListBuilder>(pool, dim_data_);
+  value_data_ = std::make_shared<PrimitiveBuilder<T>>(pool, dtype);
+  values_ = std::make_shared<ListBuilder>(pool, value_data_);
+  auto dims_field = std::make_shared<Field>("dims", dims_->type());
+  auto values_field = std::make_shared<Field>("data", values_->type());
+  auto type = std::make_shared<StructType>(std::vector<FieldPtr>({dims_field, values_field}));
+  tensors_ = std::make_shared<StructBuilder>(pool, type, std::vector<std::shared_ptr<ArrayBuilder>>({dims_, values_}));
+};
+
+template<typename T>
+Status TensorBuilder<T>::Append(const std::vector<int64_t>& dims, const elem_type* data) {
+  RETURN_NOT_OK(tensors_->Append());
+  RETURN_NOT_OK(dims_->Append());
+  RETURN_NOT_OK(values_->Append());
+  int32_t size = 1;
+  for (auto dim : dims) {
+    size *= dim;
+    RETURN_NOT_OK(dim_data_->Append(dim));
+  }
+  RETURN_NOT_OK(value_data_->Append(data, size));
+  return Status::OK(); // tensors_->Append();
+}
+
+template<typename T>
+std::shared_ptr<Array> TensorBuilder<T>::Finish() {
+  return tensors_->Finish();
+}
+
+template class TensorBuilder<UInt8Type>;
+template class TensorBuilder<Int8Type>;
+template class TensorBuilder<UInt16Type>;
+template class TensorBuilder<Int16Type>;
+template class TensorBuilder<UInt32Type>;
+template class TensorBuilder<Int32Type>;
+template class TensorBuilder<UInt64Type>;
+template class TensorBuilder<Int64Type>;
+template class TensorBuilder<FloatType>;
+template class TensorBuilder<DoubleType>;
+
+}
diff --git a/cpp/src/numbuf/tensor.h b/cpp/src/numbuf/tensor.h
new file mode 100644
index 000000000..351892758
--- /dev/null
+++ b/cpp/src/numbuf/tensor.h
@@ -0,0 +1,67 @@
+#ifndef NUMBUF_TENSOR_H
+#define NUMBUF_TENSOR_H
+
+#include <memory>
+#include <arrow/type.h>
+#include <arrow/api.h>
+
+namespace numbuf {
+
+/*! This is a class for building a dataframe where each row corresponds to
+    a Tensor (= multidimensional array) of numerical data. There are two
+    columns, "dims" which contains an array of dimensions for each Tensor
+    and "data" which contains data buffer of the Tensor as a flattened array.
+*/
+template<typename T>
+class TensorBuilder {
+public:
+  typedef typename T::c_type elem_type;
+
+  TensorBuilder(const arrow::TypePtr& dtype, arrow::MemoryPool* pool = nullptr);
+
+  /*! Append a new tensor.
+
+      \param dims
+        The dimensions of the Tensor
+
+      \param data
+        Pointer to the beginning of the data buffer of the Tensor. The
+        total length of the buffer is sizeof(elem_type) * product of dims[i] over i
+  */
+  arrow::Status Append(const std::vector<int64_t>& dims, const elem_type* data);
+
+  //! Convert the tensors to an Arrow StructArray
+  std::shared_ptr<arrow::Array> Finish();
+
+  //! Number of tensors in the column
+  int32_t length() {
+    return tensors_->length();
+  }
+
+  const arrow::TypePtr& type() {
+    return tensors_->type();
+  }
+
+private:
+	arrow::TypePtr dtype_;
+  std::shared_ptr<arrow::Int64Builder> dim_data_;
+  std::shared_ptr<arrow::ListBuilder> dims_;
+  std::shared_ptr<arrow::PrimitiveBuilder<T>> value_data_;
+  std::shared_ptr<arrow::ListBuilder> values_;
+  std::shared_ptr<arrow::StructBuilder> tensors_;
+};
+
+typedef TensorBuilder<arrow::UInt8Type> UInt8TensorBuilder;
+typedef TensorBuilder<arrow::Int8Type> Int8TensorBuilder;
+typedef TensorBuilder<arrow::UInt16Type> UInt16TensorBuilder;
+typedef TensorBuilder<arrow::Int16Type> Int16TensorBuilder;
+typedef TensorBuilder<arrow::UInt32Type> UInt32TensorBuilder;
+typedef TensorBuilder<arrow::Int32Type> Int32TensorBuilder;
+typedef TensorBuilder<arrow::UInt64Type> UInt64TensorBuilder;
+typedef TensorBuilder<arrow::Int64Type> Int64TensorBuilder;
+typedef TensorBuilder<arrow::FloatType> FloatTensorBuilder;
+typedef TensorBuilder<arrow::DoubleType> DoubleTensorBuilder;
+
+}
+
+#endif // NUMBUF_TENSOR_H
diff --git a/python/src/pynumbuf/adapters/numpy.cc b/python/src/pynumbuf/adapters/numpy.cc
new file mode 100644
index 000000000..6ca6753ca
--- /dev/null
+++ b/python/src/pynumbuf/adapters/numpy.cc
@@ -0,0 +1,88 @@
+#include "numpy.h"
+
+#include <numbuf/tensor.h>
+
+using namespace arrow;
+
+namespace numbuf {
+
+#define ARROW_TYPE_TO_NUMPY_CASE(TYPE) \
+  case Type::TYPE:                     \
+    return NPY_##TYPE;
+
+#define DESERIALIZE_ARRAY_CASE(TYPE, ArrayType, type)                        \
+  case Type::TYPE: {                                                         \
+    auto values = std::dynamic_pointer_cast<ArrayType>(content->values());   \
+    DCHECK(values);                                                          \
+    type* data = const_cast<type*>(values->raw_data())                       \
+                   + content->offset(offset);                                \
+    *out = PyArray_SimpleNewFromData(num_dims, dim.data(), NPY_##TYPE,       \
+                                       reinterpret_cast<void*>(data));       \
+  }                                                                          \
+  return Status::OK();
+
+Status DeserializeArray(std::shared_ptr<Array> array, int32_t offset, PyObject** out) {
+  DCHECK(array);
+  auto tensor = std::dynamic_pointer_cast<StructArray>(array);
+  DCHECK(tensor);
+  auto dims = std::dynamic_pointer_cast<ListArray>(tensor->field(0));
+  auto content = std::dynamic_pointer_cast<ListArray>(tensor->field(1));
+  npy_intp num_dims = dims->value_length(offset);
+  std::vector<npy_intp> dim(num_dims);
+  for (int i = dims->offset(offset); i < dims->offset(offset+1); ++i) {
+    dim[i - dims->offset(offset)] =
+      std::dynamic_pointer_cast<Int64Array>(dims->values())->Value(i);
+  }
+  switch (content->value_type()->type) {
+    DESERIALIZE_ARRAY_CASE(INT8, Int8Array, int8_t)
+    DESERIALIZE_ARRAY_CASE(INT16, Int16Array, int16_t)
+    DESERIALIZE_ARRAY_CASE(INT32, Int32Array, int32_t)
+    DESERIALIZE_ARRAY_CASE(INT64, Int64Array, int64_t)
+    DESERIALIZE_ARRAY_CASE(UINT8, UInt8Array, uint8_t)
+    DESERIALIZE_ARRAY_CASE(UINT16, UInt16Array, uint16_t)
+    DESERIALIZE_ARRAY_CASE(UINT32, UInt32Array, uint32_t)
+    DESERIALIZE_ARRAY_CASE(UINT64, UInt64Array, uint64_t)
+    DESERIALIZE_ARRAY_CASE(FLOAT, FloatArray, float)
+    DESERIALIZE_ARRAY_CASE(DOUBLE, DoubleArray, double)
+    default:
+      DCHECK(false) << "arrow type not recognized: " << content->value_type()->type;
+  }
+  return Status::OK();
+}
+
+Status SerializeArray(PyArrayObject* array, SequenceBuilder& builder) {
+  size_t ndim = PyArray_NDIM(array);
+  int dtype = PyArray_TYPE(array);
+  std::vector<int64_t> dims(ndim);
+  for (int i = 0; i < ndim; ++i) {
+    dims[i] = PyArray_DIM(array, i);
+  }
+  auto data = PyArray_DATA(array);
+  switch (dtype) {
+    case NPY_UINT8:
+      return builder.Append(dims, reinterpret_cast<uint8_t*>(data));
+    case NPY_INT8:
+      return builder.Append(dims, reinterpret_cast<int8_t*>(data));
+    case NPY_UINT16:
+      return builder.Append(dims, reinterpret_cast<uint16_t*>(data));
+    case NPY_INT16:
+      return builder.Append(dims, reinterpret_cast<int16_t*>(data));
+    case NPY_UINT32:
+      return builder.Append(dims, reinterpret_cast<uint32_t*>(data));
+    case NPY_INT32:
+      return builder.Append(dims, reinterpret_cast<int32_t*>(data));
+    case NPY_UINT64:
+      return builder.Append(dims, reinterpret_cast<uint64_t*>(data));
+    case NPY_INT64:
+      return builder.Append(dims, reinterpret_cast<int64_t*>(data));
+    case NPY_FLOAT:
+      return builder.Append(dims, reinterpret_cast<float*>(data));
+    case NPY_DOUBLE:
+      return builder.Append(dims, reinterpret_cast<double*>(data));
+    default:
+      DCHECK(false) << "numpy data type not recognized: " << dtype;
+  }
+  return Status::OK();
+}
+
+}
diff --git a/python/src/pynumbuf/adapters/numpy.h b/python/src/pynumbuf/adapters/numpy.h
new file mode 100644
index 000000000..8b8b3859a
--- /dev/null
+++ b/python/src/pynumbuf/adapters/numpy.h
@@ -0,0 +1,22 @@
+#ifndef PYNUMBUF_NUMPY_H
+#define PYNUMBUF_NUMPY_H
+
+#include <arrow/api.h>
+#include <Python.h>
+
+#define NPY_NO_DEPRECATED_API NPY_1_7_API_VERSION
+#define NO_IMPORT_ARRAY
+#define PY_ARRAY_UNIQUE_SYMBOL NUMBUF_ARRAY_API
+#include <numpy/arrayobject.h>
+
+#include <numbuf/tensor.h>
+#include <numbuf/sequence.h>
+
+namespace numbuf {
+
+arrow::Status SerializeArray(PyArrayObject* array, SequenceBuilder& builder);
+arrow::Status DeserializeArray(std::shared_ptr<arrow::Array> array, int32_t offset, PyObject** out);
+
+}
+
+#endif
diff --git a/python/src/pynumbuf/adapters/python.cc b/python/src/pynumbuf/adapters/python.cc
new file mode 100644
index 000000000..751ebf91b
--- /dev/null
+++ b/python/src/pynumbuf/adapters/python.cc
@@ -0,0 +1,193 @@
+#include "python.h"
+
+#include <sstream>
+
+#include "scalars.h"
+
+using namespace arrow;
+
+namespace numbuf {
+
+PyObject* get_value(ArrayPtr arr, int32_t index, int32_t type) {
+  PyObject* result;
+  switch (arr->type()->type) {
+    case Type::BOOL:
+      return PyBool_FromLong(std::static_pointer_cast<BooleanArray>(arr)->Value(index));
+    case Type::INT64:
+      return PyInt_FromLong(std::static_pointer_cast<Int64Array>(arr)->Value(index));
+    case Type::STRING: {
+      int32_t nchars;
+      const uint8_t* str = std::static_pointer_cast<StringArray>(arr)->GetValue(index, &nchars);
+      return PyString_FromStringAndSize(reinterpret_cast<const char*>(str), nchars);
+    }
+    case Type::FLOAT:
+      return PyFloat_FromDouble(std::static_pointer_cast<FloatArray>(arr)->Value(index));
+    case Type::DOUBLE:
+      return PyFloat_FromDouble(std::static_pointer_cast<DoubleArray>(arr)->Value(index));
+    case Type::STRUCT: {
+      auto s = std::static_pointer_cast<StructArray>(arr);
+      auto l = std::static_pointer_cast<ListArray>(s->field(0));
+      if (s->type()->child(0)->name == "list") {
+        ARROW_CHECK_OK(DeserializeList(l->values(), l->value_offset(index), l->value_offset(index+1), &result));
+      } else if (s->type()->child(0)->name == "tuple") {
+        ARROW_CHECK_OK(DeserializeTuple(l->values(), l->value_offset(index), l->value_offset(index+1), &result));
+      } else if (s->type()->child(0)->name == "dict") {
+        ARROW_CHECK_OK(DeserializeDict(l->values(), l->value_offset(index), l->value_offset(index+1), &result));
+      } else {
+        ARROW_CHECK_OK(DeserializeArray(arr, index, &result));
+      }
+      return result;
+    }
+    default:
+      DCHECK(false) << "union tag not recognized " << type;
+  }
+  return NULL;
+}
+
+Status append(PyObject* elem, SequenceBuilder& builder,
+              std::vector<PyObject*>& sublists,
+              std::vector<PyObject*>& subtuples,
+              std::vector<PyObject*>& subdicts) {
+  // The bool case must precede the int case (PyInt_Check passes for bools)
+  if (PyBool_Check(elem)) {
+    RETURN_NOT_OK(builder.Append(elem == Py_True));
+  } else if (PyFloat_Check(elem)) {
+    RETURN_NOT_OK(builder.Append(PyFloat_AS_DOUBLE(elem)));
+  } else if (PyLong_Check(elem)) {
+    int overflow = 0;
+    int64_t data = PyLong_AsLongLongAndOverflow(elem, &overflow);
+    RETURN_NOT_OK(builder.Append(data));
+    if(overflow) {
+      return Status::NotImplemented("long overflow");
+    }
+  } else if (PyInt_Check(elem)) {
+    RETURN_NOT_OK(builder.Append(PyInt_AS_LONG(elem)));
+  } else if (PyString_Check(elem)) {
+    RETURN_NOT_OK(builder.Append(PyString_AS_STRING(elem), PyString_GET_SIZE(elem)));
+  } else if (PyList_Check(elem)) {
+    builder.AppendList(PyList_Size(elem));
+    sublists.push_back(elem);
+  } else if (PyDict_Check(elem)) {
+    builder.AppendDict(PyDict_Size(elem));
+    subdicts.push_back(elem);
+  } else if (PyTuple_Check(elem)) {
+    builder.AppendTuple(PyTuple_Size(elem));
+    subtuples.push_back(elem);
+  } else if (PyArray_IsScalar(elem, Generic)) {
+    RETURN_NOT_OK(AppendScalar(elem, builder));
+  } else if (PyArray_Check(elem)) {
+    RETURN_NOT_OK(SerializeArray((PyArrayObject*) elem, builder));
+  } else if (elem == Py_None) {
+    RETURN_NOT_OK(builder.Append());
+  } else {
+    std::stringstream ss;
+    ss << "data type of " << PyString_AS_STRING(PyObject_Repr(elem))
+       << " not recognized";
+    return Status::NotImplemented(ss.str());
+  }
+  return Status::OK();
+}
+
+Status SerializeSequences(std::vector<PyObject*> sequences, std::shared_ptr<Array>* out) {
+  DCHECK(out);
+  SequenceBuilder builder(nullptr);
+  std::vector<PyObject*> sublists, subtuples, subdicts;
+  for (const auto& sequence : sequences) {
+    PyObject* item;
+    PyObject* iterator = PyObject_GetIter(sequence);
+    while (item = PyIter_Next(iterator)) {
+      RETURN_NOT_OK(append(item, builder, sublists, subtuples, subdicts));
+      Py_DECREF(item);
+    }
+    Py_DECREF(iterator);
+  }
+  std::shared_ptr<Array> list;
+  if (sublists.size() > 0) {
+    RETURN_NOT_OK(SerializeSequences(sublists, &list));
+  }
+  std::shared_ptr<Array> tuple;
+  if (subtuples.size() > 0) {
+    RETURN_NOT_OK(SerializeSequences(subtuples, &tuple));
+  }
+  std::shared_ptr<Array> dict;
+  if (subdicts.size() > 0) {
+    RETURN_NOT_OK(SerializeDict(subdicts, &dict));
+  }
+  *out = builder.Finish(list, tuple, dict);
+  return Status::OK();
+}
+
+#define DESERIALIZE_SEQUENCE(CREATE, SET_ITEM)                                \
+  auto data = std::dynamic_pointer_cast<DenseUnionArray>(array);              \
+  int32_t size = array->length();                                             \
+  PyObject* result = CREATE(stop_idx - start_idx);                            \
+  auto types = std::make_shared<Int8Array>(size, data->types());              \
+  auto offsets = std::make_shared<Int32Array>(size, data->offset_buf());      \
+  for (size_t i = start_idx; i < stop_idx; ++i) {                             \
+    if (data->IsNull(i)) {                                                    \
+      Py_INCREF(Py_None);                                                     \
+      SET_ITEM(result, i-start_idx, Py_None);                                 \
+    } else {                                                                  \
+      int32_t offset = offsets->Value(i);                                     \
+      int8_t type = types->Value(i);                                          \
+      ArrayPtr arr = data->child(type);                                       \
+      SET_ITEM(result, i-start_idx, get_value(arr, offset, type));            \
+    }                                                                         \
+  }                                                                           \
+  *out = result;                                                              \
+  return Status::OK();
+
+Status DeserializeList(std::shared_ptr<Array> array, int32_t start_idx, int32_t stop_idx, PyObject** out) {
+  DESERIALIZE_SEQUENCE(PyList_New, PyList_SetItem)
+}
+
+Status DeserializeTuple(std::shared_ptr<Array> array, int32_t start_idx, int32_t stop_idx, PyObject** out) {
+  DESERIALIZE_SEQUENCE(PyTuple_New, PyTuple_SetItem)
+}
+
+Status SerializeDict(std::vector<PyObject*> dicts, std::shared_ptr<Array>* out) {
+  DictBuilder result;
+  std::vector<PyObject*> sublists, subtuples, subdicts, dummy;
+  for (const auto& dict : dicts) {
+    PyObject *key, *value;
+    Py_ssize_t pos = 0;
+    while (PyDict_Next(dict, &pos, &key, &value)) {
+      RETURN_NOT_OK(append(key, result.keys(), dummy, dummy, dummy));
+      DCHECK(dummy.size() == 0);
+      RETURN_NOT_OK(append(value, result.vals(), sublists, subtuples, subdicts));
+    }
+  }
+  std::shared_ptr<Array> val_list;
+  if (sublists.size() > 0) {
+    RETURN_NOT_OK(SerializeSequences(sublists, &val_list));
+  }
+  std::shared_ptr<Array> val_tuples;
+  if (subtuples.size() > 0) {
+    RETURN_NOT_OK(SerializeSequences(subtuples, &val_tuples));
+  }
+  std::shared_ptr<Array> val_dict;
+  if (subdicts.size() > 0) {
+    RETURN_NOT_OK(SerializeDict(subdicts, &val_dict));
+  }
+  *out = result.Finish(val_list, val_tuples, val_dict);
+  return Status::OK();
+}
+
+Status DeserializeDict(std::shared_ptr<Array> array, int32_t start_idx, int32_t stop_idx, PyObject** out) {
+  auto data = std::dynamic_pointer_cast<StructArray>(array);
+  // TODO(pcm): error handling, get rid of the temporary copy of the list
+  PyObject *keys, *vals;
+  PyObject* result = PyDict_New();
+  ARROW_RETURN_NOT_OK(DeserializeList(data->field(0), start_idx, stop_idx, &keys));
+  ARROW_RETURN_NOT_OK(DeserializeList(data->field(1), start_idx, stop_idx, &vals));
+  for (size_t i = start_idx; i < stop_idx; ++i) {
+    PyDict_SetItem(result, PyList_GetItem(keys, i - start_idx), PyList_GetItem(vals, i - start_idx));
+  }
+  Py_XDECREF(keys); // PyList_GetItem(keys, ...) incremented the reference count
+  Py_XDECREF(vals); // PyList_GetItem(vals, ...) incremented the reference count
+  *out = result;
+  return Status::OK();
+}
+
+
+}
diff --git a/python/src/pynumbuf/adapters/python.h b/python/src/pynumbuf/adapters/python.h
new file mode 100644
index 000000000..1fa07f920
--- /dev/null
+++ b/python/src/pynumbuf/adapters/python.h
@@ -0,0 +1,22 @@
+#ifndef PYNUMBUF_PYTHON_H
+#define PYNUMBUF_PYTHON_H
+
+#include <Python.h>
+
+#include <arrow/api.h>
+#include <numbuf/dict.h>
+#include <numbuf/sequence.h>
+
+#include "numpy.h"
+
+namespace numbuf {
+
+arrow::Status SerializeSequences(std::vector<PyObject*> sequences, std::shared_ptr<arrow::Array>* out);
+arrow::Status SerializeDict(std::vector<PyObject*> dicts, std::shared_ptr<arrow::Array>* out);
+arrow::Status DeserializeList(std::shared_ptr<arrow::Array> array, int32_t start_idx, int32_t stop_idx, PyObject** out);
+arrow::Status DeserializeTuple(std::shared_ptr<arrow::Array> array, int32_t start_idx, int32_t stop_idx, PyObject** out);
+arrow::Status DeserializeDict(std::shared_ptr<arrow::Array> array, int32_t start_idx, int32_t stop_idx, PyObject** out);
+
+}
+
+#endif
diff --git a/python/src/pynumbuf/adapters/scalars.h b/python/src/pynumbuf/adapters/scalars.h
new file mode 100644
index 000000000..928289252
--- /dev/null
+++ b/python/src/pynumbuf/adapters/scalars.h
@@ -0,0 +1,54 @@
+#ifndef PYNUMBUF_SCALARS_H
+#define PYNUMBUF_SCALARS_H
+
+#include <arrow/api.h>
+
+#include <Python.h>
+#define NPY_NO_DEPRECATED_API NPY_1_7_API_VERSION
+#define NO_IMPORT_ARRAY
+#define PY_ARRAY_UNIQUE_SYMBOL NUMBUF_ARRAY_API
+#include <numpy/arrayobject.h>
+#include <numpy/arrayscalars.h>
+
+#include <numbuf/sequence.h>
+
+namespace numbuf {
+
+arrow::Status AppendScalar(PyObject* obj, SequenceBuilder& builder) {
+  if (PyArray_IsScalar(obj, Bool)) {
+    return builder.Append(((PyBoolScalarObject *)obj)->obval != 0);
+  } else if (PyArray_IsScalar(obj, Float)) {
+    return builder.Append(((PyFloatScalarObject *)obj)->obval);
+  } else if (PyArray_IsScalar(obj, Double)) {
+    return builder.Append(((PyDoubleScalarObject *)obj)->obval);
+  }
+  int64_t value = 0;
+  if (PyArray_IsScalar(obj, Byte)) {
+    value = ((PyByteScalarObject *)obj)->obval;
+  } else if (PyArray_IsScalar(obj, UByte)) {
+    value = ((PyUByteScalarObject *)obj)->obval;
+  } else if (PyArray_IsScalar(obj, Short)) {
+    value = ((PyShortScalarObject *)obj)->obval;
+  } else if (PyArray_IsScalar(obj, UShort)) {
+    value = ((PyUShortScalarObject *)obj)->obval;
+  } else if (PyArray_IsScalar(obj, Int)) {
+    value = ((PyIntScalarObject *)obj)->obval;
+  } else if (PyArray_IsScalar(obj, UInt)) {
+    value = ((PyUIntScalarObject *)obj)->obval;
+  } else if (PyArray_IsScalar(obj, Long)) {
+    value = ((PyLongScalarObject *)obj)->obval;
+  } else if (PyArray_IsScalar(obj, ULong)) {
+    value = ((PyULongScalarObject *)obj)->obval;
+  } else if (PyArray_IsScalar(obj, LongLong)) {
+    value = ((PyLongLongScalarObject *)obj)->obval;
+  } else if (PyArray_IsScalar(obj, ULongLong)) {
+    value = ((PyULongLongScalarObject *)obj)->obval;
+  } else {
+    DCHECK(false) << "scalar type not recognized";
+  }
+  return builder.Append(value);
+}
+
+} // namespace
+
+#endif // PYNUMBUF_SCALARS_H
diff --git a/python/src/pynumbuf/memory.h b/python/src/pynumbuf/memory.h
new file mode 100644
index 000000000..ebaa22c51
--- /dev/null
+++ b/python/src/pynumbuf/memory.h
@@ -0,0 +1,47 @@
+#ifndef PYNUMBUF_MEMORY_H
+#define PYNUMBUF_MEMORY_H
+
+#include <arrow/ipc/memory.h>
+
+namespace numbuf {
+
+class BufferSource : public arrow::ipc::MemorySource {
+ public:
+  virtual ~BufferSource() {}
+
+  explicit BufferSource(uint8_t* data, int64_t nbytes)
+    : data_(data), size_(nbytes) {}
+
+  arrow::Status ReadAt(int64_t position, int64_t nbytes,
+                       std::shared_ptr<arrow::Buffer>* out) override {
+    DCHECK(out);
+    DCHECK(position + nbytes <= size_) << "position: " << position << " nbytes: " << nbytes << "size: " << size_;
+    *out = std::make_shared<arrow::Buffer>(data_ + position, nbytes);
+    return arrow::Status::OK();
+  }
+
+  arrow::Status Close() override {
+    return arrow::Status::OK();
+  }
+
+  arrow::Status Write(int64_t position, const uint8_t* data,
+                      int64_t nbytes) override {
+    DCHECK(position >= 0 && position < size_);
+    DCHECK(position + nbytes <= size_) << "position: " << position << " nbytes: " << nbytes << "size: " << size_;
+    uint8_t* dst = data_ + position;
+    memcpy(dst, data, nbytes);
+    return arrow::Status::OK();
+  }
+
+  int64_t Size() const override {
+    return size_;
+  }
+
+private:
+ uint8_t* data_;
+ int64_t size_;
+};
+
+} // namespace
+
+#endif // PYNUMBUF_MEMORY_H
diff --git a/python/src/pynumbuf/numbuf.cc b/python/src/pynumbuf/numbuf.cc
new file mode 100644
index 000000000..6b4c524dd
--- /dev/null
+++ b/python/src/pynumbuf/numbuf.cc
@@ -0,0 +1,205 @@
+#include <Python.h>
+#include <arrow/api.h>
+#include <arrow/ipc/memory.h>
+#include <arrow/ipc/adapter.h>
+#define NPY_NO_DEPRECATED_API NPY_1_7_API_VERSION
+#define PY_ARRAY_UNIQUE_SYMBOL NUMBUF_ARRAY_API
+#include <numpy/arrayobject.h>
+
+#include <iostream>
+
+#include <arrow/ipc/metadata.h>
+
+#include "adapters/python.h"
+#include "memory.h"
+
+using namespace arrow;
+using namespace numbuf;
+
+extern "C" {
+
+// Error handling
+
+static PyObject *NumbufError;
+
+int PyObjectToArrow(PyObject* object, std::shared_ptr<Array> **result) {
+  if (PyCapsule_IsValid(object, "arrow")) {
+    *result = reinterpret_cast<std::shared_ptr<Array>*>(PyCapsule_GetPointer(object, "arrow"));
+    return 1;
+  } else {
+    PyErr_SetString(PyExc_TypeError, "must be an 'arrow' capsule");
+    return 0;
+  }
+}
+
+static void ArrowCapsule_Destructor(PyObject* capsule) {
+  delete reinterpret_cast<std::shared_ptr<Array>*>(PyCapsule_GetPointer(capsule, "arrow"));
+}
+
+std::shared_ptr<RowBatch> make_row_batch(std::shared_ptr<Array> data) {
+  auto field = std::make_shared<Field>("list", data->type());
+  std::shared_ptr<Schema> schema(new Schema({field}));
+  return std::shared_ptr<RowBatch>(new RowBatch(schema, data->length(), {data}));
+}
+
+/*! Serializes a Python list into an Arrow array.
+
+    \param args
+      The argument must be a Python list
+
+    \returns
+      A Python "arrow" capsule containing the arrow::Array
+*/
+PyObject* serialize_list(PyObject* self, PyObject* args) {
+  PyObject* value;
+  if (!PyArg_ParseTuple(args, "O", &value)) {
+    return NULL;
+  }
+  std::shared_ptr<Array>* result = new std::shared_ptr<Array>();
+  if (PyList_Check(value)) {
+    Status s = SerializeSequences(std::vector<PyObject*>({value}), result);
+    if (!s.ok()) {
+      PyErr_SetString(NumbufError, s.ToString().c_str());
+      return NULL;
+    }
+    return PyCapsule_New(reinterpret_cast<void*>(result), "arrow", &ArrowCapsule_Destructor);
+  }
+  return NULL;
+}
+
+/*! Number of bytes the serialized version of the object will take.
+
+    \param args
+      A Python "arrow" capsule containing the arrow::Array
+
+    \returns
+      Size of the object in memory once it is serialized
+*/
+PyObject* get_serialized_size(PyObject* self, PyObject* args) {
+  std::shared_ptr<Array>* data;
+  if (!PyArg_ParseTuple(args, "O&", &PyObjectToArrow, &data)) {
+    return NULL;
+  }
+  auto batch = make_row_batch(*data);
+  int64_t size = 0;
+  ARROW_CHECK_OK(arrow::ipc::GetRowBatchSize(batch.get(), &size));
+  return PyInt_FromLong(size);
+}
+
+/*! Serialize an arrow::Array into a buffer.
+
+    \param args
+      A Python "arrow" capsule containing the arrow::Array and
+      a python memoryview object the data will be written to
+
+    \return
+      The arrow metadata offset for the arrow metadata
+*/
+PyObject* write_to_buffer(PyObject* self, PyObject* args) {
+  std::shared_ptr<Array>* data;
+  PyObject* memoryview;
+  if (!PyArg_ParseTuple(args, "O&O", &PyObjectToArrow, &data, &memoryview)) {
+    return NULL;
+  }
+  if (!PyMemoryView_Check(memoryview)) {
+    return NULL;
+  }
+  auto batch = make_row_batch(*data);
+  Py_buffer* buffer = PyMemoryView_GET_BUFFER(memoryview);
+  auto target = std::make_shared<BufferSource>(reinterpret_cast<uint8_t*>(buffer->buf), buffer->len);
+  int64_t metadata_offset;
+  ARROW_CHECK_OK(ipc::WriteRowBatch(target.get(), batch.get(), 0, &metadata_offset));
+  return PyInt_FromLong(metadata_offset);
+}
+
+/*! Serialize schema metadata associated to and arrow::Array
+
+    \param args
+      A Python "arrow" capsule containing the arrow::Array
+
+    \return
+      A bytearray object containing the schema metadata
+*/
+PyObject* get_schema_metadata(PyObject* self, PyObject* args) {
+  std::shared_ptr<Array>* data;
+  if (!PyArg_ParseTuple(args, "O&", &PyObjectToArrow, &data)) {
+    return NULL;
+  }
+  auto batch = make_row_batch(*data);
+  std::shared_ptr<Buffer> buffer;
+  ARROW_CHECK_OK(ipc::WriteSchema(batch->schema().get(), &buffer));
+  auto ptr = reinterpret_cast<const char*>(buffer->data());
+  return PyByteArray_FromStringAndSize(ptr, buffer->size());
+}
+
+/*! Read serialized data from buffer and produce an arrow capsule
+
+    \param args
+      A Python memoryview from which data will be loaded,
+      a Python bytearray containing the metadata and the metadata_offset
+
+    \return
+      A Python "arrow" capsule containing the arrow data
+*/
+PyObject* read_from_buffer(PyObject* self, PyObject* args) {
+  PyObject* memoryview;
+  PyObject* metadata;
+  int64_t metadata_offset;
+  if (!PyArg_ParseTuple(args, "OOl", &memoryview, &metadata, &metadata_offset)) {
+    return NULL;
+  }
+
+  auto ptr = reinterpret_cast<uint8_t*>(PyByteArray_AsString(metadata));
+  auto schema_buffer = std::make_shared<Buffer>(ptr, PyByteArray_Size(metadata));
+  std::shared_ptr<ipc::Message> message;
+  ARROW_CHECK_OK(ipc::Message::Open(schema_buffer, &message));
+  DCHECK_EQ(ipc::Message::SCHEMA, message->type());
+  std::shared_ptr<ipc::SchemaMessage> schema_msg = message->GetSchema();
+  std::shared_ptr<Schema> schema;
+  ARROW_CHECK_OK(schema_msg->GetSchema(&schema));
+
+  Py_buffer* buffer = PyMemoryView_GET_BUFFER(memoryview);
+  auto source = std::make_shared<BufferSource>(reinterpret_cast<uint8_t*>(buffer->buf), buffer->len);
+  std::shared_ptr<arrow::ipc::RowBatchReader> reader;
+  ARROW_CHECK_OK(arrow::ipc::RowBatchReader::Open(source.get(), metadata_offset, &reader));
+  std::shared_ptr<arrow::RowBatch> data;
+  ARROW_CHECK_OK(reader->GetRowBatch(schema, &data));
+
+  std::shared_ptr<Array>* result = new std::shared_ptr<Array>();
+  *result = data->column(0);
+  return PyCapsule_New(reinterpret_cast<void*>(result), "arrow", &ArrowCapsule_Destructor);
+}
+
+/*!
+*/
+PyObject* deserialize_list(PyObject* self, PyObject* args) {
+  std::shared_ptr<Array>* data;
+  if (!PyArg_ParseTuple(args, "O&", &PyObjectToArrow, &data)) {
+    return NULL;
+  }
+  PyObject* result;
+  ARROW_CHECK_OK(DeserializeList(*data, 0, (*data)->length(), &result));
+  return result;
+}
+
+static PyMethodDef NumbufMethods[] = {
+ { "serialize_list", serialize_list, METH_VARARGS, "serialize a Python list" },
+ { "deserialize_list", deserialize_list, METH_VARARGS, "deserialize a Python list" },
+ { "get_serialized_size", get_serialized_size, METH_VARARGS, "get the number of bytes the object will occupy once serialized" },
+ { "write_to_buffer", write_to_buffer, METH_VARARGS, "write serialized data to buffer"},
+ { "read_from_buffer", read_from_buffer, METH_VARARGS, "read serialized data from buffer"},
+ { "get_schema_metadata", get_schema_metadata, METH_VARARGS, "return the schema of an arrow object"},
+ { NULL, NULL, 0, NULL }
+};
+
+PyMODINIT_FUNC initlibnumbuf(void) {
+  PyObject* m;
+  m = Py_InitModule3("libnumbuf", NumbufMethods, "Python C Extension for Numbuf");
+  char numbuf_error[] = "numbuf.error";
+  NumbufError = PyErr_NewException(numbuf_error, NULL, NULL);
+  Py_INCREF(NumbufError);
+  PyModule_AddObject(m, "numbuf_error", NumbufError);
+  import_array();
+}
+
+}
diff --git a/python/test/runtest.py b/python/test/runtest.py
new file mode 100644
index 000000000..f3697d51f
--- /dev/null
+++ b/python/test/runtest.py
@@ -0,0 +1,63 @@
+import unittest
+import libnumbuf
+import numpy as np
+from numpy.testing import assert_equal
+
+TEST_OBJECTS = [[1, "hello", 3.0], 42, 43L, "hello world", 42.0, 1L << 62,
+                (1.0, "hi"), None, (None, None), ("hello", None),
+                True, False, (True, False),
+                {True: "hello", False: "world"},
+                {"hello" : "world", 1: 42, 1.0: 45}, {},
+                np.int8(3), np.int32(4), np.int64(5),
+                np.uint8(3), np.uint32(4), np.uint64(5),
+                np.float32(1.0), np.float64(1.0)]
+
+class SerializationTests(unittest.TestCase):
+
+  def roundTripTest(self, data):
+    serialized = libnumbuf.serialize_list(data)
+    result = libnumbuf.deserialize_list(serialized)
+    assert_equal(data, result)
+
+  def testSimple(self):
+    self.roundTripTest([1, 2, 3])
+    self.roundTripTest([1.0, 2.0, 3.0])
+    self.roundTripTest(['hello', 'world'])
+    self.roundTripTest([1, 'hello', 1.0])
+    self.roundTripTest([{'hello': 1.0, 'world': 42}])
+    self.roundTripTest([True, False])
+
+  def testNone(self):
+    self.roundTripTest([1, 2, None, 3])
+
+  def testNested(self):
+    self.roundTripTest([{"hello": {"world": (1, 2, 3)}}])
+    self.roundTripTest([((1,), (1, 2, 3, (4, 5, 6), "string"))])
+    self.roundTripTest([{"hello": [1, 2, 3]}])
+    self.roundTripTest([{"hello": [1, [2, 3]]}])
+    self.roundTripTest([{"hello": (None, 2, [3, 4])}])
+    self.roundTripTest([{"hello": (None, 2, [3, 4], np.ndarray([1.0, 2.0, 3.0]))}])
+
+  def numpyTest(self, t):
+    a = np.random.randint(0, 10, size=(100, 100)).astype(t)
+    self.roundTripTest([a])
+
+  def testArrays(self):
+    for t in ["int8", "uint8", "int16", "uint16", "int32", "uint32", "float32", "float64"]:
+      self.numpyTest(t)
+
+  def testRay(self):
+    for obj in TEST_OBJECTS:
+      self.roundTripTest([obj])
+
+  def testBuffer(self):
+    for (i, obj) in enumerate(TEST_OBJECTS):
+      x = libnumbuf.serialize_list([1, 2, 3])
+      schema = libnumbuf.get_schema_metadata(x)
+      size = libnumbuf.get_serialized_size(x) + 4096 # INITIAL_METADATA_SIZE in arrow
+      buff = np.zeros(size, dtype="uint8")
+      metadata_offset = libnumbuf.write_to_buffer(x, memoryview(buff))
+      array = libnumbuf.read_from_buffer(memoryview(buff), schema, metadata_offset)
+
+if __name__ == "__main__":
+    unittest.main()