Work
A handful of things I've made, and a few credentials along the way.
Projects
Real-time Virtual Human Streaming System
A real-time (25 FPS) virtual human assistant streaming system that integrates LLM, TTS, and video generation to deliver interactive, low-latency conversational experiences for streaming and marketing/sales application.
- Frame rate
- 25 FPS
- Cold start
- -60%
- Built an end-to-end multi-thread pipeline to integrate multiple services that helps the system achieves stable real-time performance.
- Designed a specific module-level failover mechanism that reduces the unplanned downtime of system from minutes to 2 seconds each week.
- Optimized resource usage 75% and reduced cold start by 60% by implementing the streaming's graph algorithm and caching strategy.
- Integrated Gitlab CI/CD to automate the deployment process, reducing manual intervention and improve overall system reliability.
- Deployed the system on AWS that achieves scalable and cost-effective infrastructure.
- Redis
- RabbitMQ
- LiveKit
- RTMP
- Docker
- CI/CD
- AWS
AI-powered News Video Generation System
An AI-powered news video generation pipeline based on talking face generation and video composition that automates the production workflow for news platforms, reducing video production time from days to hours. The system enables scalable generation of up to 10 news videos per day while significantly reducing manual video editing costs through automated scripting, voice generation, media processing, and video composition.
- Throughput
- 10/day
- Storage cost
- -60%
- Researched SOTA talking face generation and video composition solutions and integrated them as the base of the system.
- Built and end-to-end news video generation pipeline that helps the system generate up to 10 news videos per day.
- Resolved limited resource usage by using divide-to-conquer mechanism which split videos into multiple parts that helps the system generate up-to-hour news videos.
- Designed the asynchronous video I/O mechanism that optimized around 40% of pipeline speed and 60% storage cost.
- PyTorch
- TensorRT
- Docker
Scalable Resume Tailoring & Job Application System
An AI-powered resume tailoring application that eliminates the need to manually customize resumes for each job application by enabling batch resume tailoring and large-scale job application workflows.
- Inference time
- -60%
- Cost / message
- -30%
- Built the full-stack resume builder application that helps tailoring, managing resumes and job applications.
- Implemented a multi-agent workflow (criticizing-before-writing) that improves the tailoring system's quality and reliability.
- Reduced the 60% of inference time and 30% of the cost each message by calling parallelly tools and using prompt engineering techniques.
- Designed an expert-based evaluation pipeline for effective qualifying the multi-agent system.
- OpenAI
- FastAPI
- PostgreSQL
- LLM
Lecture Video Generation for Online Teaching
An application that automates the creation of online lectures through speech and video generation technologies, helping instructors eliminate time-consuming video editing workflows and concentrate on delivering engaging learning content.
- Uptime
- 99.9%
- Idle cost
- 0
- Built a lecture video generation pipeline using Text-To-Speech and AI-driven video generation techniques that reduces the time of lecture video production from days to hours.
- Designed a serverless architecture that helps achieving zero idle-cost and 99.9% uptime (based on AWS and RunPod Serverless).
- Implemented an effective queue processing mechanism for traffic spikes that protects the limited GPU resources and improves the overall system's resiliency.
- Stable Diffusion
- Text-To-Speech
- AWS
- Serverless
- Docker
Certificates
AWS Certified Solutions Architect – Associate (SAA-C03)
AWS · 12/2025
TOEIC Listening & Reading — Score 830
IIG · 7/2024
Paper presented at the National Science Conference FAIR'2023
FAIR'2023 · 7/2023