https://www.luckma.io/architecture daily 0.75 2020-07-09 https://images.squarespace-cdn.com/content/v1/5f024410d9611817b20d4cf7/1594110013723-5YR08666V5LYE8O5QQPP/luckma_architecture_img_only.png Architecture https://images.squarespace-cdn.com/content/v1/5f024410d9611817b20d4cf7/1594011937485-8FUCK25XP5PJPV4QRKPO/apache_beam_overview.png Architecture - Data Pipeline Framework Our data pipeline leverages open source framework called Apache Beam. Our data pipeline utilizes Apache Beam’s batching processing capability to process training data in parallel within multi-threaded context. Our feature scaling component prepares data for live inference request via gRPC/REST protocol or further scales feature vectors in the data pipeline for data transformation . https://images.squarespace-cdn.com/content/v1/5f024410d9611817b20d4cf7/1594012764229-2NHQU4K5DHF6JLZLBTCT/apache_spark.jpg Architecture - Data Transformation Our Data transformation engine is implemented using Apache Spark. Our engine spawns Apache spark master and worker clusters, launches JVM to start Apache PySpark context, creates RDD from Spark dataframe consisting of scaled feature vectors to transform datasets into Tensorflow Record. During the benchmark, data transformation using Apache Spark over TensorflowRecordWriter showed more than 500% performance increase. https://images.squarespace-cdn.com/content/v1/5f024410d9611817b20d4cf7/1594013179731-V8RL83AP1CRMSIS8M1CI/dask_logo.png Architecture - Feature Extraction Rather than relying on traditional Pandas for exploratory data analysis, our feature extraction engine extracts knowledge and insights from the data by utilizing Dask. During our benchmark, we discovered that performance gain using Dask’s apply function’s heavily depends on the usage of vectorized operations within the implemented apply function context. Within the optimal context, using Dask resulted over x100 gain over Panda’s dataframe apply. https://images.squarespace-cdn.com/content/v1/5f024410d9611817b20d4cf7/1594110194134-OZIGXLA62CKKPDDKYJYK/homomorphic_encryption.png Architecture - Data Security - Encryption Before the training data gets sent into the wire, we perform proprietary data encryption on our dataset similar to homomorphic data encryption. This reduces the vulnerability of the dataset in case of security breach. It’s important to note that our encryption algorithm doesn’t introduce overhead on training pipeline’s model training because it requires no decryption for the training algorithm. https://images.squarespace-cdn.com/content/v1/5f024410d9611817b20d4cf7/1594022860384-HKJ558M24TEFY784V5MA/sagemaker_pipemode.jpg Architecture - AWS Sagemaker Pipe Mode AWS Sagemaker loads the model training data from AWS S3. There is significant performance difference between File Mode and Pipe Mode. In File mode, the training data is downloaded first to an encrypted EBS volume attached to the training instance prior to training. In contrast, data is streamed directly to the training algorithm while it is running in Pipe mode. Our model training pipeline utilizes AWS’s pipe mode via the usage of PipeModeDataset to reduce AWS S3 cost for the file transfer between AWS S3 and our instance. As shown in the benchmark graph, Pipe mode reduces the job start up time by x10 and x2 increase in I/O throughput. https://images.squarespace-cdn.com/content/v1/5f024410d9611817b20d4cf7/1594024830017-PYY9ITXQO7UMGIJIX6VJ/tf_optmized.png Architecture - Tensorflow Best Practices Our model training pipeline’s input_fn implementation follows Tensorflow’s best practices outlined in the Tensorflow’s documentation. Use the prefetch transformation to overlap the work of a producer and consumer. Parallelize the data reading transformation using the interleave transformation. Parallelize the map transformation by setting the num_parallel_calls argument. Use the cache transformation to cache data in memory during the first epoch Vectorize user-defined functions passed in to the map transformation Reduce memory usage when applying the interleave, prefetch, and shuffle transformations. https://images.squarespace-cdn.com/content/v1/5f024410d9611817b20d4cf7/1594070614482-DUS8E3ZU356BAZSO5I4Y/tf_docker.png Architecture - Dockerized Inference Pipeline Our inference pipeline serving hundreds of Tensorflow Serving model artifacts has been implemented with our proprietary docker-compose/protobuf configuration manager. Inference pipeline detects changes in the model artifacts, manages configuration changes, promotes the best performing Tensorflow Serving models, and updates Docker Container without any downtime in the inference pipeline. https://images.squarespace-cdn.com/content/v1/5f024410d9611817b20d4cf7/1594027884349-DYTLVBZP9V5HBVRLTV7V/tfx.png Architecture - Tensorflow Serving (TFX) We leverage Tensorflow Serving for our inference pipeline. The generic TFX library is compiled with no CPU or GPU optimization. In order to enable GPU inference capability, we compiled Tensorflow binary using Bazel and deployed it on the Docker container for the machine learning model inference. https://www.luckma.io/infrastructure daily 0.75 2020-07-09 https://images.squarespace-cdn.com/content/v1/5f024410d9611817b20d4cf7/1594064159908-T7837UV2BA5TWIIM28ZN/aws_sagemaker.jpeg Infrastructure https://images.squarespace-cdn.com/content/v1/5f024410d9611817b20d4cf7/1594068757563-SOERJN3UWLUI8INJU0YP/aws_sagemaker_docker.png Infrastructure - Docker with AWS Sagemaker A single file called Dockerfile defines your Infrastructure as Code (IaS) that specifies an environment, configuration, and access for your application. When you compile Dockerfile, you produce a Docker image which gets uploaded to AWS Elastic Container Registry (ECR). AWS Sagemaker launches Docker Container by pulling your Docker Image from AWS ECR then begins your machine learning model training at your AWS EC2 instance. Once the model training is complete, model artifacts are uploaded to AWS S3. https://images.squarespace-cdn.com/content/v1/5f024410d9611817b20d4cf7/1594064696729-W4SKJRETG7U6TOQHSLA7/nvidia_gpu.jpg Infrastructure - Nvidia GPU Core mathematical operations performed in deep learning are suitable to be parallelized. Parallelization capacities in Graphical Processing Units (GPU) are higher than CPUs, because GPUs have far more cores than Central Processing Units (CPUs). Benchmark shows GPU is 4-5 times faster than CPU, according to the tests performed on GPU server and CPU server. Nvidia offers GPUs for the machine learning. These GPUs are directly accessible for the machine learning within AWS EC2 p/g instances via AWS deep learning AMIs. https://images.squarespace-cdn.com/content/v1/5f024410d9611817b20d4cf7/1594066047675-XZQ4HSPRSS5Q510L7MEZ/cuda.png Infrastructure - Nvidia CUDA Nvidia CUDA is GPU accelerated libraries for the Nvidia Graphical Processing Units (GPU). Nvidia CUDA has the direct deep learning support for the open source deep learning frameworks such as Tensorflow or Pytorch. Open source framework such as Tensorflow has CUDA version dependency. It’s important to meet the requirements for your infrastructure to operate as expected. https://images.squarespace-cdn.com/content/v1/5f024410d9611817b20d4cf7/1594066985441-INLWKG9JBIY97NE6PCU5/Aro%2BHa_0393.jpg Infrastructure - Nvidia Docker Our Dockerized inference pipeline is closely integrated with nvidia-docker which provides GPU pass-through of CUDA driver between our Host OS and Docker container. With nvidia-docker, one can utilize GPUs to train machine learning models within Docker container which is agnostic to Host OS and embrace Infrastructure as Code (IaS) design pattern. https://www.luckma.io/technology daily 0.75 2022-01-04 https://images.squarespace-cdn.com/content/v1/5f024410d9611817b20d4cf7/1641277834232-RYFPLTLMNYMTBXHPAVAW/reinforcement_learning.gif Technology - Make it stand out Whatever it is, the way you tell your story online can make all the difference. https://images.squarespace-cdn.com/content/v1/5f024410d9611817b20d4cf7/1594154368208-703SP5RNNKTCUO7LN04E/overfitting.png Technology - Orthogonal Optimization One of challenges in machine learning is to find optimal set of hyperparameters to tune your model while preventing your models from overfitting. Rather than employing naive regularization technique such as early stopping which stops the training entirely if no improvement over specified steps, we attempt to apply orthogonal model tuning optimization. In context of model tuning, the orthogonal optimization works as follows: fit training set well on the cost function, fit the evaluation set well on the cost function fit the testing set well on the cost function perform well on real world data. Early stopping isn’t orthogonal because it both fits the training set less well and improves the dev set performance at the same time. https://images.squarespace-cdn.com/content/v1/5f024410d9611817b20d4cf7/1594162182601-JWJSRSR7PPM89ET2R21R/htune.jpg Technology - Automated Hyperparameter Tuning Due to large quantity of models we manage, it doesn’t scale to manually fine tune our models one by one. We utilize the power of AWS Cloud Infrastructure with AWS Sagemaker to automate hyperparameter tuning. By defining a range of hyperparameters to tune within AWS Sagemaker enables us to find optimal set of hyperparameters in parallel by spawning multiple AWS EC2 instances. https://images.squarespace-cdn.com/content/v1/5f024410d9611817b20d4cf7/1594152243815-1H727M97TGO02DSJRKPF/rnn_structure.png Technology - Long Short Term Memory (LSTM) We utilize Recurrent Neural Network to interpret sequential time series data. By utilizing forget gates within Long Short Term Term Memory implementation, we prevent back propagated errors from vanishing or exploding. This implementation of RNN enables us to learn tasks that can potentially require memory of thousands or maybe millions steps earlier. https://images.squarespace-cdn.com/content/v1/5f024410d9611817b20d4cf7/1641278034136-VT3L3DQSUVXKPVG1DWV5/mmdp.png Technology - Markov Decision Process We engineer our reinforcement learning environment within Markov Decision Process where we define the environment, states, and actions an AI agent can take. The goal of reinforcement learning is to maximize cumulative reward as much as possible. https://www.luckma.io/home daily 1.0 2025-05-13 https://images.squarespace-cdn.com/content/v1/5f024410d9611817b20d4cf7/1641279023962-GQJPSZDAYP4AG4BCKFCP/reinforcement_learning.gif Home - Make it stand out Whatever it is, the way you tell your story online can make all the difference. https://images.squarespace-cdn.com/content/v1/5f024410d9611817b20d4cf7/1593998738736-8PUTZ44UKHG2SZV9H0IS/aws_dark3.png Home - AWS Cloud Infrastructure Our software infrastructure requires scalable, reliable, computing infrastructure with enhanced security while minimizing the latency. AWS offers various machine learning products to meet our requirement. https://images.squarespace-cdn.com/content/v1/5f024410d9611817b20d4cf7/1594000939933-V342EQGLSP19ZBB88HJC/aws_architecture.png Home - AWS Cloud System Architecture Our data engine ingests time-series big data from hundreds of distributed data sources and produces millions of data points daily. We leverage the latest machine learning technologies in AWS cloud infrastructure to engineer our data into the decision by AI. https://images.squarespace-cdn.com/content/v1/5f024410d9611817b20d4cf7/1594001016966-R1TVX03DTKP73Y3EI1ZE/machine_learning_tech.jpg Home - Machine Learning Technology Machine learning involves computers learning from the data in supervised and/or unsupervised ways to execute tasks. We leverage various open source software frameworks to engineer our products. https://www.luckma.io/our-team daily 0.75 2025-05-13 https://images.squarespace-cdn.com/content/v1/5f024410d9611817b20d4cf7/5e8405ff-fa16-494f-92fa-10e53502de34/1671699395714.jpg Our Team - Ja Hon Sim Ja Sim (JS) is an entrepreneur, reverse engineer, white hacker, certified enterprise software architect and full stack engineer in Python, Java, AWS Machine Learning, AWS Big Data. He worked at fortune global IT companies such as Google, Intuit, Salesforce, and Yahoo as software engineer in Silicon Valley for over 10 years. He self taught Machine Learning, AI, Big Data, AWS, Google Cloud, and Android. He graduated from the University of Colorado at Boulder with B.S. degree in computer science and mathematics with 4.0 semester GPA. He was enrolled into the National Scholars Honor Society for his achievement. He completed many M.S. courses for the self-challenge while he was an undergraduate student. He worked as a research assistant for Prof. Rick Han’s wireless sensor node operating system called Mantis OS. He had also established multiple start up businesses and self taught various software technologies such as Machine Learning and Android during his spare time. You can download his resume below or contact him via his LinkedIn or via email - jae.lim@luckma.io https://images.squarespace-cdn.com/content/v1/5f024410d9611817b20d4cf7/1594355981954-XIRZ9WOOW2HJEZX5QM9X/cert_machine_learning.png Our Team Abilities Validated by the Certification Select and justify the appropriate ML approach for a given business problem Identify appropriate AWS services to implement ML solutions Design and implement scalable, cost-optimized, reliable, and secure ML solutions Recommended Knowledge and Experience Experience developing, architecting, or running ML/deep learning in AWS Cloud The ability to express the intuition behind basic ML algorithms Experience performing basic hyperparameter optimization Experience with ML and deep learning frameworks The ability to follow model-training best practices The ability to follow deployment and operational best practices https://images.squarespace-cdn.com/content/v1/5f024410d9611817b20d4cf7/1594356021102-U59ILUQO33LDIANSEN88/cert_python.png Our Team Abilities Validated by the Certification PCAP – Certified Associate in Python Programming certification shows that the individual is familiar with general computer programming concepts like conditional execution, loops, Python programming language syntax, semantics, and the runtime environment, as well as with general coding techniques and object-oriented programming. Becoming PCAP certified ensures that the individual is fully acquainted with all the primary means provided by Python 3 to enable her/him to start her/his own studies, and to open a path to the developer’s career. https://images.squarespace-cdn.com/content/v1/5f024410d9611817b20d4cf7/1594355996469-V489XXXQGJ96Q6E5Y03P/cert_big_data.png Our Team Abilities Validated by the Certification Implement core AWS Big Data services according to architecture best practices Design and maintain Big Data Leverage tools to automate data analysis Recommended Knowledge and Experience Candidates to hold an AWS Certified Cloud Practitioner or a current Associate-level certification: AWS Certified Solutions Architect - Associate, AWS Certified Developer - Associate or AWS Certified SysOps Administrator - Associate Background in defining and architecting AWS Big Data services with the ability to explain how they fit in the data life cycle of collection, ingestion, storage, processing, and visualization Minimum five years hands-on experience in a data analytics field Experience in designing a scalable and cost-effective architecture to process data https://images.squarespace-cdn.com/content/v1/5f024410d9611817b20d4cf7/1594355890888-2A275AXMRACSK8J7NRZU/aws_cert_tier.png Our Team https://images.squarespace-cdn.com/content/v1/5f024410d9611817b20d4cf7/1594356050102-NMTQ9SIC21WSOUIYX3VO/cert_java.png Our Team Abilities Validated by the Certification OOPs - Constuctors, Abstract Classes, Interfaces, File I/O, Serialization Collections – List, Map, Set, Generics, Checked/Unchecked Exceptions Java Keywords – Static, Final, volatile, synchronized, transient, this, super JVM and Memory Management, Multithreading and Synchronization Network Protocols - IP, HTTP, TCP, FTP, UDP, Sockets, RMI, RCP JSP / Servlets / Struts / Spring / Dependency Injection (IoC) Service Oriented Architecture / Web Services – SOAP / REST HTML, CSS, Javascript and JQuery, XML and JSON Frameworks - Swing, SWT, AWT, JavaFX, Java Applets SQL Queries – Inner Outer Joins, Group By, Triggers, Stored Procedures https://www.luckma.io/contact daily 0.75 2020-09-10