diff --git a/.gitignore b/.gitignore
index b067edd..55d0f3f 100644
--- a/.gitignore
+++ b/.gitignore
@@ -1 +1,3 @@
 /Manifest.toml
+
+example.tfrecord
\ No newline at end of file
diff --git a/.travis.yml b/.travis.yml
index f3bd2ac..3570473 100644
--- a/.travis.yml
+++ b/.travis.yml
@@ -3,7 +3,7 @@ language: julia
 notifications:
   email: false
 julia:
-  - 1.0
+  - 1.3
   - 1.5
   - nightly
 os:
diff --git a/Project.toml b/Project.toml
index 3f987e8..3d09c53 100644
--- a/Project.toml
+++ b/Project.toml
@@ -3,8 +3,21 @@ uuid = "841416d8-1a6a-485a-b0fc-1328d0f53d5e"
 authors = ["Jun Tian <tianjun.cpp@gmail.com> and contributors"]
 version = "0.1.0"
 
+[deps]
+BufferedStreams = "e1450e63-4bb3-523b-b2a4-4ffa8c0fd77d"
+CRC32c = "8bf52ea8-c179-5cab-976a-9e18b702a9bc"
+CodecZlib = "944b1d66-785c-5afd-91f1-9de20f533193"
+Glob = "c27321d9-0574-5035-807b-f59d2c89b15c"
+MacroTools = "1914dd2f-81c6-5fcd-8719-6d5c9610ff09"
+ProtoBuf = "3349acd9-ac6a-5e09-bcdb-63829b23a429"
+
 [compat]
-julia = "1"
+julia = "1.3"
+BufferedStreams = "1.0"
+CodecZlib = "0.7"
+Glob = "1.3"
+MacroTools = "0.5"
+ProtoBuf = "0.8"
 
 [extras]
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
diff --git a/README.md b/README.md
index e2d1635..488b09a 100644
--- a/README.md
+++ b/README.md
@@ -1,3 +1,55 @@
 # TFRecord
 
 [![Build Status](https://travis-ci.com/JuliaReinforcementLearning/TFRecord.jl.svg?branch=master)](https://travis-ci.com/JuliaReinforcementLearning/TFRecord.jl)
+
+## Usage
+
+### Install
+
+```julia
+julia> ] add TFRecord
+```
+
+### Write TFRecord
+
+```julia
+using TFRecord
+
+writer = TFRecordWriter("example.tfrecord")
+
+for i in 1:10
+    write(writer, Dict(
+        "feature1" => rand(Bool),
+        "feature2" => rand(1:5),
+        "feature3" => rand(("cat", "dog", "chicken", "horse", "goat")),
+        "feature4" => randn(Float32),
+    ))
+end
+
+close(writer)
+```
+
+Here we write `10` observations into the file `example.tfrecord`. Internally each dictionary is converted into a `TFRecord.Example` first, which is a known prototype by TensorFlow. Note that the type of key must be `AbstractString` and the type of value can be one of the following types:
+
+- `Bool`, `Int64`, `Float32`, `AbstractString`
+- `Vector` of the above types
+
+For customized data types, you need to convert it into `TFRecord.Example` first.
+
+### Read TFRecord
+
+```julia
+reader = TFRecordReader("example.tfrecord")
+
+for example in reader
+    println(example)
+end
+```
+
+For more fine-grained control, please read the doc:
+
+```julia
+julia> ? TFRecordReader
+
+julia> ? TFRecordWriter
+```
diff --git a/src/TFRecord.jl b/src/TFRecord.jl
index 10dceb5..66d3dee 100644
--- a/src/TFRecord.jl
+++ b/src/TFRecord.jl
@@ -1,5 +1,6 @@
 module TFRecord
 
-# Write your package code here.
+include("jlout/example_pb.jl")
+include("core.jl")
 
 end
diff --git a/src/core.jl b/src/core.jl
new file mode 100644
index 0000000..46f5fc2
--- /dev/null
+++ b/src/core.jl
@@ -0,0 +1,171 @@
+export TFRecordReader, TFRecordWriter
+
+using CRC32c
+using Glob
+using Base.Threads
+using CodecZlib
+using BufferedStreams
+using MacroTools: @forward
+
+# Ref: https://github.com/tensorflow/tensorflow/blob/295ad2781683835be974faba0a191528d8079768/tensorflow/core/lib/hash/crc32c.h#L50-L59
+
+const MASK_DELTA = 0xa282ead8
+
+mask(crc::UInt32) = ((crc >> 15) | (crc << 17)) + MASK_DELTA
+
+function unmask(masked_crc::UInt32)
+    rot = masked_crc - MASK_DELTA
+    ((rot >> 17) | (rot << 15))
+end
+
+"""
+
+Ref: https://github.com/tensorflow/tensorflow/blob/295ad2781683835be974faba0a191528d8079768/tensorflow/core/lib/io/record_reader.cc#L164-L199
+
+Each record is stored in the following format:
+
+```
+uint64 n
+uint32 masked_crc32_of_n
+byte   data[n]
+uint32 masked_crc32_of_data
+```
+"""
+function read_record(io::IO)
+    n = read(io, sizeof(UInt64))
+    masked_crc32_n = read(io, UInt32)
+    crc32c(n) == unmask(masked_crc32_n) || error("record corrupted, did you set the correct compression?")
+
+    data = read(io, Int(reinterpret(UInt64, n)[]))  # !!! watch https://github.com/JuliaIO/TranscodingStreams.jl/pull/104
+    masked_crc32_data = read(io, UInt32)
+    crc32c(data) == unmask(masked_crc32_data) || error("record corrupted, did you set the correct compression?")
+    data
+end
+
+#####
+# TFRecordReader
+#####
+
+"""
+    TFRecordReader(s;kwargs...)
+
+# Keyword Arguments
+
+- `compression=nothing`. No compression by default. Optional values are `:zlib` and `:gzip`.
+- `bufsize=1024*1024`. Set the buffer size of internal `BufferedOutputStream`. The default value is `1M`. Suggested value is between `1M`~`100M`.
+- `channel_size=1000`. The number of pre-fetched elements.
+
+!!!note
+
+    To enable reading records from multiple files concurrently, remember to set the number of threads correctly (See [JULIA_NUM_THREADS](https://docs.julialang.org/en/v1/manual/environment-variables/#JULIA_NUM_THREADS)). Unfortunately, the feature is currently broken. Please watch https://github.com/JuliaIO/ProtoBuf.jl/issues/140 .
+
+"""
+struct TFRecordReader{T}
+    ch::Channel
+end
+
+@forward TFRecordReader.ch Base.close, Base.iterate, Base.isopen, Base.take!
+
+TFRecordReader(s::String;kwargs...) = TFRecordReader{Example}(identity, s;kwargs...)
+
+function TFRecordReader{T}(f, s;compression=nothing,bufsize=1024*1024, channel_size=1000) where T
+    files = glob(s)
+    length(files) > 0 || error("can not find any files under: $s")
+    chnl = Channel{T}(channel_size) do ch
+        # watch https://github.com/JuliaIO/ProtoBuf.jl/issues/140
+        #= @threads =# for file_name in files
+            open(file_name, "r") do io
+
+                io = BufferedInputStream(io, bufsize)
+                if compression == :gzip
+                    io = GzipDecompressorStream(io)
+                elseif compression == :zlib
+                    io = ZlibDecompressorStream(io)
+                else
+                    isnothing(compression) || throw(ArgumentError("unsupported compression method: $compression"))
+                end
+
+                while !eof(io)
+                    example = readproto(IOBuffer(read_record(io)), Example())
+                    put!(ch, f(example))
+                end
+            end
+        end
+    end
+    TFRecordReader{T}(chnl)
+end
+
+#####
+# TFRecordWriter
+#####
+
+struct TFRecordWriter{X<:IO}
+    io::X
+end
+
+"""
+    TFRecordWriter(s;compression=nothing, bufsize=1024*1024)
+
+Supported `compression` methods are: `:gzip` or `:zlib`.
+Default value is `nothing`, which means do not do compression.
+`bufsize` is used to set the size of buffer used by an internal
+`BufferedOutputStream`, the default value is `1M` (1024*1024).
+You may want to change it to a larger value when writing large datasets,
+for example `100M`.
+"""
+function TFRecordWriter(s::AbstractString;compression=nothing, bufsize=1024*1024)
+    io = BufferedOutputStream(open(s, "w"), bufsize)
+    if compression == :gzip
+        io = GzipCompressorStream(io)
+    elseif compression == :zlib
+        io = ZlibCompressorStream(io)
+    else
+        isnothing(compression) || throw(ArgumentError("unsupported compression method: $compression"))
+    end
+    TFRecordWriter(io)
+end
+
+Base.close(w::TFRecordWriter) = close(w.io)
+
+Base.write(w::TFRecordWriter, x) = write(w, convert(Example, x))
+
+function Base.write(w::TFRecordWriter, x::Example)
+    buff = IOBuffer()
+    writeproto(buff, x)
+
+    data_crc = mask(crc32c(seekstart(buff)))
+    data = take!(seekstart(buff))
+    n = length(data)
+
+    buff = IOBuffer()
+    write(buff, n)
+    n_crc = mask(crc32c(seekstart(buff)))
+
+    write(w.io, n)
+    write(w.io, n_crc)
+    write(w.io, data)
+    write(w.io, data_crc)
+end
+
+#####
+# convert
+#####
+
+Base.convert(::Type{Feature}, x::Int) = Feature(;int64_list=Int64List(value=[x]))
+Base.convert(::Type{Feature}, x::Bool) = Feature(;int64_list=Int64List(value=[Int(x)]))
+Base.convert(::Type{Feature}, x::Float32) = Feature(;float_list=FloatList(value=[x]))
+Base.convert(::Type{Feature}, x::AbstractString) = Feature(;bytes_list=BytesList(value=[unsafe_wrap(Vector{UInt8}, x)]))
+
+Base.convert(::Type{Feature}, x::Vector{Int}) = Feature(;int64_list=Int64List(value=x))
+Base.convert(::Type{Feature}, x::Vector{Bool}) = Feature(;int64_list=Int64List(value=convert(Vector{Int}, x)))
+Base.convert(::Type{Feature}, x::Vector{Float32}) = Feature(;float_list=FloatList(value=x))
+Base.convert(::Type{Feature}, x::Vector{<:AbstractString}) = Feature(;bytes_list=BytesList(value=[unsafe_wrap(Vector{UInt8}, s) for s in x]))
+Base.convert(::Type{Feature}, x::Vector{Array{UInt8,1}}) = Feature(;bytes_list=BytesList(value=x))
+
+Base.convert(::Type{Features}, x::Dict) = Features(;feature=Dict(k=>convert(Feature, v) for (k, v) in x))
+
+function Base.convert(::Type{Example}, x::Dict)
+    d = Example()
+    d.features = convert(Features, x)
+    d
+end
diff --git a/src/jlout/example_pb.jl b/src/jlout/example_pb.jl
new file mode 100644
index 0000000..f599f6b
--- /dev/null
+++ b/src/jlout/example_pb.jl
@@ -0,0 +1,73 @@
+# syntax: proto3
+using ProtoBuf
+import ProtoBuf.meta
+
+mutable struct BytesList <: ProtoType
+    value::Base.Vector{Array{UInt8,1}}
+    BytesList(; kwargs...) = (o=new(); fillunset(o); isempty(kwargs) || ProtoBuf._protobuild(o, kwargs); o)
+end #mutable struct BytesList
+
+mutable struct FloatList <: ProtoType
+    value::Base.Vector{Float32}
+    FloatList(; kwargs...) = (o=new(); fillunset(o); isempty(kwargs) || ProtoBuf._protobuild(o, kwargs); o)
+end #mutable struct FloatList
+const __pack_FloatList = Symbol[:value]
+meta(t::Type{FloatList}) = meta(t, ProtoBuf.DEF_REQ, ProtoBuf.DEF_FNUM, ProtoBuf.DEF_VAL, true, __pack_FloatList, ProtoBuf.DEF_WTYPES, ProtoBuf.DEF_ONEOFS, ProtoBuf.DEF_ONEOF_NAMES, ProtoBuf.DEF_FIELD_TYPES)
+
+mutable struct Int64List <: ProtoType
+    value::Base.Vector{Int64}
+    Int64List(; kwargs...) = (o=new(); fillunset(o); isempty(kwargs) || ProtoBuf._protobuild(o, kwargs); o)
+end #mutable struct Int64List
+const __pack_Int64List = Symbol[:value]
+meta(t::Type{Int64List}) = meta(t, ProtoBuf.DEF_REQ, ProtoBuf.DEF_FNUM, ProtoBuf.DEF_VAL, true, __pack_Int64List, ProtoBuf.DEF_WTYPES, ProtoBuf.DEF_ONEOFS, ProtoBuf.DEF_ONEOF_NAMES, ProtoBuf.DEF_FIELD_TYPES)
+
+mutable struct Feature <: ProtoType
+    bytes_list::BytesList
+    float_list::FloatList
+    int64_list::Int64List
+    Feature(; kwargs...) = (o=new(); fillunset(o); isempty(kwargs) || ProtoBuf._protobuild(o, kwargs); o)
+end #mutable struct Feature
+const __oneofs_Feature = Int[1,1,1]
+const __oneof_names_Feature = [Symbol("kind")]
+meta(t::Type{Feature}) = meta(t, ProtoBuf.DEF_REQ, ProtoBuf.DEF_FNUM, ProtoBuf.DEF_VAL, true, ProtoBuf.DEF_PACK, ProtoBuf.DEF_WTYPES, __oneofs_Feature, __oneof_names_Feature, ProtoBuf.DEF_FIELD_TYPES)
+
+mutable struct Features_FeatureEntry <: ProtoType
+    key::AbstractString
+    value::Feature
+    Features_FeatureEntry(; kwargs...) = (o=new(); fillunset(o); isempty(kwargs) || ProtoBuf._protobuild(o, kwargs); o)
+end #mutable struct Features_FeatureEntry (mapentry)
+
+mutable struct Features <: ProtoType
+    feature::Base.Dict{AbstractString,Feature} # map entry
+    Features(; kwargs...) = (o=new(); fillunset(o); isempty(kwargs) || ProtoBuf._protobuild(o, kwargs); o)
+end #mutable struct Features
+
+mutable struct Example <: ProtoType
+    features::Features
+    Example(; kwargs...) = (o=new(); fillunset(o); isempty(kwargs) || ProtoBuf._protobuild(o, kwargs); o)
+end #mutable struct Example
+
+mutable struct FeatureList <: ProtoType
+    feature::Base.Vector{Feature}
+    FeatureList(; kwargs...) = (o=new(); fillunset(o); isempty(kwargs) || ProtoBuf._protobuild(o, kwargs); o)
+end #mutable struct FeatureList
+
+mutable struct FeatureLists_FeatureListEntry <: ProtoType
+    key::AbstractString
+    value::FeatureList
+    FeatureLists_FeatureListEntry(; kwargs...) = (o=new(); fillunset(o); isempty(kwargs) || ProtoBuf._protobuild(o, kwargs); o)
+end #mutable struct FeatureLists_FeatureListEntry (mapentry)
+
+mutable struct FeatureLists <: ProtoType
+    feature_list::Base.Dict{AbstractString,FeatureList} # map entry
+    FeatureLists(; kwargs...) = (o=new(); fillunset(o); isempty(kwargs) || ProtoBuf._protobuild(o, kwargs); o)
+end #mutable struct FeatureLists
+
+mutable struct SequenceExample <: ProtoType
+    context::Features
+    feature_lists::FeatureLists
+    SequenceExample(; kwargs...) = (o=new(); fillunset(o); isempty(kwargs) || ProtoBuf._protobuild(o, kwargs); o)
+end #mutable struct SequenceExample
+
+export Example, SequenceExample, BytesList, FloatList, Int64List, Feature, Features_FeatureEntry, Features, FeatureList, FeatureLists_FeatureListEntry, FeatureLists
+# mapentries: "FeatureLists_FeatureListEntry" => ("AbstractString", "FeatureList"), "Features_FeatureEntry" => ("AbstractString", "Feature")
diff --git a/src/proto/example.proto b/src/proto/example.proto
new file mode 100644
index 0000000..b770e68
--- /dev/null
+++ b/src/proto/example.proto
@@ -0,0 +1,55 @@
+// This .proto file is based on those that are in https://github.com/tensorflow/tensorflow/tree/master/tensorflow/core/example
+
+syntax = "proto3";
+option cc_enable_arenas = true;
+
+message Example
+{
+    Features features = 1;
+}
+
+message SequenceExample
+{
+    Features context = 1;
+    FeatureLists feature_lists = 2;
+}
+
+message BytesList
+{
+    repeated bytes value = 1;
+}
+
+message FloatList
+{
+    repeated float value = 1 [packed = true];
+}
+
+message Int64List
+{
+    repeated int64 value = 1 [packed = true];
+}
+
+message Feature
+{
+    oneof kind
+    {
+        BytesList bytes_list = 1;
+        FloatList float_list = 2;
+        Int64List int64_list = 3;
+    }
+}
+
+message Features
+{
+    map<string, Feature> feature = 1;
+}
+
+message FeatureList
+{
+    repeated Feature feature = 1;
+}
+
+message FeatureLists
+{
+    map<string, FeatureList> feature_list = 1;
+}
diff --git a/test/runtests.jl b/test/runtests.jl
index 733185f..918f212 100644
--- a/test/runtests.jl
+++ b/test/runtests.jl
@@ -2,5 +2,32 @@ using TFRecord
 using Test
 
 @testset "TFRecord.jl" begin
-    # Write your tests here.
+
+    writer = TFRecordWriter("example.tfrecord")
+
+    n = 10
+    f1 = rand(Bool, n)
+    f2 = rand(1:5, n)
+    f3 = rand(("cat", "dog", "chicken", "horse", "goat"), n)
+    f4 = rand(Float32, n)
+
+    for i in 1:n
+        write(writer, Dict(
+            "feature1" => f1[i],
+            "feature2" => f2[i],
+            "feature3" => f3[i],
+            "feature4" => f4[i],
+        ))
+    end
+
+    close(writer)
+
+    reader = TFRecordReader("example.tfrecord")
+
+    for (i, example) in enumerate(reader)
+        @test example.features.feature["feature1"].int64_list.value[] == Int(f1[i])
+        @test example.features.feature["feature2"].int64_list.value[] == f2[i]
+        @test String(example.features.feature["feature3"].bytes_list.value[]) == f3[i]
+        @test example.features.feature["feature4"].float_list.value[] == f4[i]
+    end
 end