Amr Amer
Download CV

Amr Amer

Machine Learning Engineer | Computer Vision | Deep Learning

Machine Learning Engineer with an M.Sc. in Visual Computing from Saarland University and experience at DFKI (German Research Center for Artificial Intelligence). I build production-ready ML systems end-to-end, from versioned, quality-checked data pipelines and distributed GPU training to multimodal generative models, high-performance C++ inference and deployment, with a focus on computer vision.

On the data and MLOps side, I build reproducible pipelines with DVC, Airflow, Spark and MLflow, add automated data-quality gates, and ship models with Docker, Kubernetes and CI/CD on GitHub Actions.

Currently seeking roles in Machine Learning Engineering and Computer Vision.

Email LinkedIn GitHub Download CV
Open to work +49 162 6667294
Portrait of Amr Amer

What I do

Master's thesis

Personality-Aware Non-verbal Behavior Generation in Dyadic Interactions

Multimodal generative model for generating realistic listener avatars in dyadic conversations, conditioned on personality traits. The model predicts facial expressions, head motion and upper-body gestures of a listener from the speaker's audio and motion signals.

Problem

Human conversations rely heavily on non-verbal cues such as facial expressions, eye gaze and body movement. Most prior work does not model how these behaviors vary with personality traits.

Approach

Developed a multimodal generative architecture combining:

  • Transformer encoder for temporal modeling
  • VQ-VAE for quantized motion representation learning
  • SMPL-X body parameters for realistic face and upper-body avatar generation
  • Integration of audio, facial motion and body gesture signals

The model generates personality-aware listener behavior sequences in response to a speaker.

Results

Face FID
6.15
Face P-FID
10.31
Upper-body FID
43.16
Upper-body P-FID
87.73

User study: participants correctly identified extroverted vs. introverted avatars in 86% of cases, and 71% preferred the personality-aware model over the personality-agnostic baseline.

Tech stack

  • PyTorch
  • Transformers
  • VQ-VAE
  • SMPL-X (PIXIE)
  • Librosa
  • OpenCV
  • CUDA
  • TensorBoard
  • SLURM
  • Enroot
  • Multi-GPU training (A100)

Computer vision system

Badminton-VisionAI: AI Match Analytics from Video

Badminton-VisionAI is an end-to-end computer vision pipeline that converts badminton match videos into structured performance analytics for players and coaches. The system tracks both players and the shuttlecock, detects shot events, projects motion onto a mini-court representation and generates a downloadable coach-style performance report.

Tracking and on-court intelligence. Player tracking, shuttle tracking, shot detection, court mapping.
Analytics dashboard (Streamlit). Player statistics, heatmaps, shot analysis, coach-ready report.

Technical challenges

Badminton tracking presents several challenges for computer vision systems:

  • Extremely small and fast shuttlecock
  • Frequent player occlusions
  • Rapid direction changes that break naive trackers

This project combines object detection, multi-object tracking, geometric projection and temporal analytics to produce coach-readable match insights, not just raw detections.

Key features

  • Stable multi-object player tracking
  • Shuttlecock trajectory tracking with speed estimation
  • Shot type classification with temporal stabilization
  • Mini-court projection using homography
  • Interactive analytics dashboard for match analysis
  • Automated coach report with movement and shot statistics

Tech stack

  • Python
  • OpenCV
  • YOLO
  • Multi-object tracking
  • Court homography
  • Trajectory projection
  • Shot detection
  • Temporal label stabilization
  • Streamlit
  • Plotly
  • ReportLab PDF export

This project reflects my interest in real-world computer vision systems that transform model outputs into actionable insights.

Data pipelines and MLOps

AV Sensor DataOps: From Raw Sensor Logs to a Monitored 3D Detection API

Automated pipeline that turns raw autonomous-vehicle sensor logs (nuScenes camera and LiDAR) into versioned, quality-checked, ML-ready datasets, then trains a camera detector, fuses it with LiDAR for 3D object detection, evaluates it nuScenes-style and serves it as a monitored API.

  1. Raw logsDVC
  2. BronzeParquet
  3. Silversynced, features
  4. Quality gateschema, sensor, label checks
  5. Goldtags, splits, labels
  6. TrainYOLOv8, MLflow
  7. Evaluateper-scenario slices
  8. Promotion gateregression checks
  9. ServeONNX, FastAPI
Every stage is a DVC stage with hashed inputs and outputs. Airflow triggers training when a new dataset version lands, and Prometheus and Grafana monitor data and model health.

What it does

  • Versioning and lineage: every dataset and model is reproducible from a Git tag with DVC; MLflow runs carry the commit and dataset hash
  • Data-quality gate: Pandera schemas plus sensor checks (sync offsets, dropped frames, corrupt files, degraded LiDAR, suspect calibration) and label checks; errors fail the pipeline
  • Orchestration: Airflow with data-aware scheduling, so a new dataset version triggers training, evaluation and promotion
  • Camera + LiDAR fusion: 2D detections lifted to 3D with LiDAR frustums, evaluated with nuScenes-style mAP, ATE, ASE, AOE and NDS-lite
  • Serving and monitoring: ONNX Runtime and FastAPI with Prometheus metrics, Grafana dashboards and a latency benchmark for CPU, GPU and Jetson targets

CI/CD

GitHub Actions runs lint and tests, the full DVC pipeline on synthetic data, the promotion gate, a reproducibility check and an API smoke test on every push, then publishes the pipeline image. A scheduled workflow re-runs the data stages and opens a pull request, so dataset releases are reviewed like code.

Tech stack

  • DVC
  • Airflow
  • PySpark
  • pandas / PyArrow
  • Pandera
  • MLflow
  • YOLOv8
  • ONNX Runtime
  • FastAPI
  • Prometheus
  • Grafana
  • Docker
  • Kubernetes
  • GitHub Actions

Other projects

Computer vision, autonomous driving

Semantic Segmentation for Autonomous Vehicles

End-to-end semantic segmentation of urban street scenes for autonomous driving perception.

  • Dataset: BDD100K
  • Task: multi-class semantic segmentation
  • Classes: road, traffic light, traffic sign, vehicle, person, bicycle, background
  • Evaluation: mean Intersection over Union (mIoU)
  • Results: mIoU ≈ 0.45, with strong performance on dominant classes (road ≈ 0.88, vehicle ≈ 0.78)
  • Optimization: systematic hyperparameter optimization via experiment sweeps
  • Pipeline: data preparation → training → evaluation → deployment and visualization
  • PyTorch
  • fastai
  • Semantic segmentation
  • Hyperparameter optimization
  • Experiment tracking
GitHub repository

Medical imaging

3D Brain Tumor Segmentation (MRI)

Multi-label 3D semantic segmentation of glioma sub-regions from volumetric MRI scans.

  • Dataset: Medical Segmentation Decathlon / BraTS (multi-modal MRI)
  • Model: 3D SegResNet (MONAI)
  • Input modalities: FLAIR, T1, T1Gd, T2
  • Target structures: whole tumor (WT), tumor core (TC), enhancing tumor (ET)
  • Training and evaluation: Dice loss and mean Dice
  • Results: mean Dice = 0.78 on validation (WT 0.90, TC 0.82, ET 0.59)
  • Pipeline: preprocessing → training → inference → deployment and visualization
  • PyTorch
  • MONAI
  • 3D SegResNet
  • Multi-modal MRI
  • 3D medical image transforms
  • Sliding-window inference
  • Weights & Biases
GitHub repository

Vision-language models

Realtime Vision Captioning

Exploratory project focused on applying vision-language models in realtime webcam-based settings.

  • Realtime image captioning and visual question answering using BLIP via Hugging Face Transformers
  • Image classification using ResNet-50 pretrained on ImageNet
  • Realtime Gradio-based webcam application for live image captioning and classification with adjustable inference parameters

The project is implemented as a set of Jupyter notebooks, progressing from offline image understanding to a realtime vision application.

  • Python
  • PyTorch
  • Torchvision
  • Hugging Face Transformers
  • Gradio
  • PIL
  • NumPy
GitHub repository
AI conversational agent web interface with an avatar

Conversational AI Ongoing

AI Conversational Agent with Voice, Avatar and Web Interface

Conversational AI system enabling real-time voice and text interaction through a custom web-based interface.

  • Speech-to-text and text-to-speech for natural dialogue
  • LLM-powered conversational reasoning
  • Human-like conversational avatar integrated into the UI
  • Modular frontend-backend design for extensibility
  • Python
  • OpenAI GPT
  • Speech-to-text
  • Text-to-speech
  • HTML
  • CSS
  • JavaScript
  • Bootstrap
  • jQuery

Technical profile

Computer vision tasks
Object detection, segmentation, tracking, 3D object detection, camera and LiDAR sensor fusion, video motion analysis, digital avatar generation
Models and frameworks
PyTorch, TensorFlow, OpenCV, YOLO, SAM, CNNs, Vision Transformers (ViT), TrackNet, Supervision
Training and evaluation
Transfer learning, fine-tuning, loss design, metric selection, hyperparameter optimization, multi-GPU training (CUDA)
Data pipelines and quality
Bronze, silver and gold data layers, DVC dataset versioning and lineage, Airflow orchestration, PySpark, pandas / PyArrow, Parquet, Pandera schema and data-quality checks
Data and experimentation
Dataset preparation, data augmentation, efficient data loading, MLflow, Weights & Biases, TensorBoard, ablation studies
Deployment and inference
Docker, Kubernetes, FastAPI, ONNX Runtime, GPU inference pipelines, batch and real-time inference, AWS (EC2 GPU), Streamlit (demos)
CI/CD and monitoring
GitHub Actions, automated testing and reproducibility checks, model promotion gates, Prometheus, Grafana
Compute and infrastructure
SLURM (HPC), GPU job scheduling, multi-node training, large-scale GPU experiments
Software engineering foundations
Python, C++, object-oriented design, version control (Git), debugging, logging, unit testing

Beyond work

Outside of my professional work, I enjoy aerial photography and videography using drones. This section highlights selected visual projects created in my free time.

View more drone videos on Vimeo

Contact

If you're hiring or would like to collaborate, I'd be glad to hear from you.