- Improved the Virtual Quality Controller legacy system, elevating its audio processing time from 450 hours to an outstanding 20,000 hours in a day. This was achieved through the migration from traditional on-premise virtual machines technologies to modern cloud-based Kubernetes (specifically AWS EKS) for more efficient cluster management.- Hardened the security of the legacy system by transitioning it from a vulnerable public-facing environment to a secure internal system, mitigating potential risks and safeguarding sensitive data.- Implemented auto scaling solutions to optimize resource allocation and accommodate dynamic workloads effectively using HPA kubernetes, Keda, Prometheus. Significantly reduced provisioning time for new pods on new nodes in Kubernetes, decreasing it from 15 minutes to just 2 minutes, thereby improving the efficiency of the auto scaling process.- Collaborated closely with the accounting team and directly with the CEO and CTO to achieve a 30% reduction in infrastructure costs as a percentage of revenue. Implemented best practices in FinOps, such as Kubecost, to manage the financial aspects of cloud computing and optimizing costs in cloud environments (AWS, Azure).- Setup tooling for “Infrastructure as Code” like Helm, Terraform, Ansible or AWS CloudFormation, following the “cattle, not pets” approach to reduce operational overhead.- Used open-source technologies and tools to accomplish specific tasks in monitoring systems like Grafana, Prometheus or App/ Database open-source monitoring like Proxmox (Virtual machine management platform), PgAdmin/ mongoDB compass (Database), Redinsight (Redis), AKHQ (Kafka).- Used open-source technologies and tools to accomplish specific tasks in monitoring systems and logging system. Put logrotate in linux system or index policies in Elastic Search to prevent data loss due to data overflow.- Create a custom log processor within the EFK/ ELK stack to collect essential data, thereby avoid data overflow.

- Maintained the AI system running in production: working closely with 10+ PIC AI modules, being involved in the code review process to build a pipeline for extracting information from identity cards. Optimized processing time using concurrency techniques and applied monitoring technologies to observe service when running in Docker. - Built CI/ CD pipeline to decrease build, test and deploy process.- Mitigating flicker effect by when applying models learned from image dataset turn into video in real life, working with 5+ people to serve beauty effect plugin on OBS studio.- Participated in research to integrate AI models into production.

- Resolved problems solution in OCR systems, Voice command and Face Anti Spoofing, primarily based on public implementations of SOTA Deep Learning models- Tuned models and augmented data, evaluated the predictions, exported and served light models- Applied image processing techniques (SIFT, HoG, Hough line, Contour, Bag of words, Frequency domain) to handle inadequate inputs and outputs.

- Researched and built services for Smart Home Systems using Microcontrollers and Cloud Platforms- Documented technical reports, script instructions, then published guides to the community at tapit.vn