I build intelligent visual systems that see, understand and generate, from multimodal 3D avatar generation and real-time sports analytics to medical imaging and autonomous driving perception.

My work lives at the intersection of computer vision, deep learning and generative modeling, with a focus on pushing research-grade systems toward scalable, real-world solutions. 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.

A multimodal generative model for synthesizing 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.

Results: face FID 6.15 and P-FID 10.31, upper-body FID 43.16 and P-FID 87.73, better than a personality-agnostic baseline. In a user study, participants identified the extroverted avatar in 86% of cases.

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

Thesis website | Repository | Full thesis (PDF)

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

Tech stack: Python · PyTorch · OpenCV · YOLO · ByteTrack · TrackNet · SAM · Streamlit · Plotly · Docker · CUDA · Homography · Body pose estimation · 3D motion trajectories · CI/CD pipeline

CV pipeline video | Dashboard video | Repository | Live demo _{(free server, can take about a minute to wake up)}

An 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.

flowchart LR
  A[Raw sensor logs] --> B[Bronze] --> C[Silver] --> D{Quality gate} --> E[Gold] --> F[Train] --> G[Evaluate] --> H{Promotion gate} --> I[ONNX + FastAPI]

• Versioning and lineage: every stage is a DVC stage; datasets and models are reproducible from a Git tag, and MLflow runs carry the commit and dataset hash
• Data quality: Pandera schemas plus sensor and label checks (sync offsets, dropped frames, corrupt files, degraded LiDAR, suspect calibration) that fail the pipeline on errors
• Orchestration and CI/CD: Airflow triggers training when a new dataset version lands; GitHub Actions runs tests, the full pipeline on synthetic data and a reproducibility check on every push
• Serving and monitoring: ONNX Runtime and FastAPI with Prometheus metrics and Grafana dashboards

Tech stack: DVC · Airflow · PySpark · Pandera · MLflow · YOLOv8 · ONNX Runtime · FastAPI · Prometheus · Grafana · Docker · Kubernetes · GitHub Actions

Semantic Segmentation for Autonomous Vehicles Multi-class segmentation of urban street scenes on BDD100K, with systematic hyperparameter sweeps. mIoU ≈ 0.45 (road ≈ 0.88, vehicle ≈ 0.78). PyTorch · fastai · Weights & Biases	3D Brain Tumor Segmentation (MRI) Multi-label segmentation of glioma sub-regions from 3D multimodal MRI with a 3D SegResNet. Mean Dice 0.78 (WT 0.90, TC 0.82, ET 0.59). PyTorch · MONAI · sliding-window inference · Weights & Biases
Realtime Vision Captioning Jupyter notebooks that progress from offline image understanding to a real-time webcam app for live image captioning (BLIP) and classification (ResNet-50). PyTorch · Hugging Face Transformers · Gradio	AI Conversational Agent (in progress) Real-time voice and text interaction, LLM-powered reasoning and a digital-human interface in a custom web app. Python · OpenAI GPT · speech-to-text · text-to-speech · JavaScript

For the full overview of my work, visit amramer.github.io.