Generative AI and Computer Vision Engineer with hands-on experience delivering optimized deep learning systems in production. Currently develop and deploy MEIKTV’s face-swap, talking-face, and portrait-animation services…
Generative AI and Computer Vision Engineer with hands-on experience delivering optimized deep learning systems in production. Currently develop and deploy MEIKTV’s face-swap, talking-face, and portrait-animation services (SONIC, LivePortrait, GeneFace++), where tuning the inference stack cut paste-back processing time by 50% and roughly doubled the number of users a single GPU can serve. Previously built AI products end-to-end for freelance clients: LLM fine-tuning (LoRA/QLoRA), RAG and multi-agent systems, and satellite-imagery segmentation.
<p>Designed and delivered the company’s flagship face-swap product end-to-end, from model selection through production deployment: a modular pipeline chaining YOLOv8n-face detection, 2D-FAN4 landmark alignment, ArcFace R50 identity embedding, InSwapper/HyperSwap identity transfer (256–1024 px) and GFPGAN 1.4 enhancement, serving both image and video jobs with multi-person selection and per-request model switching, shipped as containerized services behind a FastAPI/WebSocket orchestration layer.</p><p> Cut paste-back processing time by 50% by re-engineering the mask-blur and inverse-affine compositing stage, and reduced per-frame latency and GPU memory by roughly 40% by converting the full model chain to ONNX/TensorRT (FP16, dynamic batching, custom plugins) with startup warm-up and reusable device buffers, allowing the same GPU fleet to serve about 2× more concurrent users at no additional hardware cost.</p><p> Eliminated false-positive swaps in production by replacing the face-only detector with a joint person + face detection model, and raised visual output quality by tracing eye-gaze instability and post-swap jitter to frame-to-frame landmark fluctuation, then benchmarking alternative landmark detectors and temporal-smoothing strategies.</p><p> Delivered the audio-driven talking-face and portrait-animation services (SONIC CVPR 2025, LivePortrait, GeneFace++), including Time-Aware Position Shift Fusion that removes identity drift and jitter in videos longer than 10 minutes at zero extra compute, unlocking full-length content the previous system could not produce; guided the team’s video-generation roadmap by benchmarking 3D causal video VAEs (Wan 2.1, SVD-XT, LeanVAE) for the best quality-to-compression trade-off.</p><p> Architected an identity-consistent character-replacement pipeline, evaluating it across three DiT-based video architectures — Wan 2.2 Animate, DreamID-V and MiniMax H3 — reframing them from generic image-to-video reenactment models into precise identity-propagation tools.</p>
<p>Video Colorization via Adapted WAN 3D VAE: Modified a pretrained 3D VAE encoder for grayscale input with custom causal padding and weight initialization. Built a training pipeline converting movies to 5-second grayscale-color paired clips and evaluated three loss strategies. Achieved PSNR 38.82 dB and SSIM 0.9885 on reconstruction validation.<br>
Mining Area Segmentation from Satellite Imagery: Developed a SegFormer (MIT-b2) system for satellite imagery, achieving 98.19% accuracy. Designed a custom ResidualAdapter to convert 12-channel satellite data to 3-channel input and implemented binary DiceLoss for improved segmentation.<br>
Real-Estate AI Phone Agent: Built a real-time voice agent using OpenAI API, Twilio, and ElevenLabs TTS with FastAPI + WebSocket for real-time interactions, handling sales, rental, and installment inquiries for real-estate clients.<br>
Multi-Agent AI Conversation System: Created a modular Flask assistant powered by Claude-3 with specialized agents. Used structured prompts, chain-of-thought reasoning, and MongoDB for persistent multi-turn conversations.<br>
Vector Database + Knowledge Graph Architecture: Engineered a unified data architecture integrating vector databases with Neo4j graph database, combining semantic search with knowledge-graph relationships for advanced data insights.<br>
LLM Fine-Tuning (LoRA/QLoRA): Fine-tuned Llama 3.1, Jais 13B (Arabic), Qwen, Silma, and Cohere with parameter-efficient fine-tuning. Built an automated training pipeline with dynamic GPU detection to auto-configure batch sizes and gradient accumulation steps.<br>
Face Restoration Model Optimization: Benchmarked 8+ face restoration models; implemented batch processing and converted models to TensorRT for enhanced performance.<br>
Coffee Fruit Maturity Detection & Counting: Developed a YOLOv8-based detection system with dual deployment: TFLite for on-device Flutter processing and a Python backend with Firebase real-time integration for remote analysis and visualization.</p>
The AI-first Recruitment Operating System that unifies and automates the entire hiring lifecycle for modern enterprises worldwide.
Dr Job FZ LLC. 2026 © All Rights Reserved
