Amir
Soltani
I build vision systems that have to work outside the lab: on a phone in a gym, on a drone above a fire line, on a Jetson in a truck cab, on a clinician's review screen.
Senior computer vision and machine learning engineer. 10+ years across medical imaging, remote sensing, driver assistance and industrial monitoring, covering the whole model lifecycle from data and annotation to deployment. I am most useful where labels are scarce, objects are small and the answer is needed in milliseconds. Lately I have been connecting vision models to language: retrieval with CLIP, DINOv2 and SigLIP, text-guided annotation and LLM pipelines for clinical documents.
The landscape behind this page is a 3D point cloud, drawn live and swept by a scan line. Scrolling flies you over it. You are seeing everything in visible light. Change the sensor with the switch at the bottom of the screen, or press 1 to 4.
Showcase
Projects with pictures, each framed the way a detector would mark it. Select a picture to enlarge it.
Experience
Six roles since 2016. Several were remote and part-time or project-based, so they ran side by side from 2022 to 2024. The bars show when. Select a role for details.
Generative AI & LLM Researcher
May 2025 to Oct 2026Nova Vision, Tallinn, Estonia. Remote, part-time.
- Built an early stroke-screening workflow that combines facial movement from video with voice features and ECG signals.
- Modelled cardiovascular risk from clinical and genomic variables with a tabular Transformer and gated fusion.
- Added CLIP, DINOv2 and SigLIP embeddings to clinical image retrieval, and evaluated Grounding DINO with SAM 3 for text-guided masks.
- Built OCR and LLM pipelines that turn unstructured clinical documents into structured, searchable records, plus retrieval-backed assistants for summaries and domain questions.
Python, PyTorch, MONAI, OpenCV, Hugging Face Transformers, CLIP, DINOv2, SigLIP, Grounding DINO, SAM 2, SAM 3, FastAPI, MLflow, Docker
Computer Vision Engineer
Jan 2023 to Feb 2025VRVis GmbH, Vienna, Austria. Remote, flexible schedule.
- Fine-tuned Prithvi-100M on Sentinel-2 time series to separate coniferous from broadleaf forest.
- Built tiled detection for small plant species and used SAHI with horizontal and oriented boxes on large satellite scenes.
- Built animal detection and counting for UAV imagery, later refining masks with YOLOv11-seg and SAM 2.
- Built thermal drone models for early fire ignition and thin smoke, reducing false alarms by 23%.
- Mentored junior ML engineers and coordinated delivery across the team.
Python, PyTorch, MMDetection, YOLOv8, SAHI, SAM 2, TorchGeo, Prithvi-100M, GDAL, Rasterio, GeoPandas, Docker
Head of ML & Computer Vision
Feb 2022 to Mar 2024Sculp AI, Texas, USA. Remote, full-time.
- Led development of AI Fitness Coach, which gives exercise-form feedback from real-time pose estimation.
- Built joint-angle tracking and repetition counting for more than 40 exercises, with on-device inference under 30 ms.
- Shipped PhysioTrack, a movement-analysis workflow for rehabilitation reviewed by licensed physiotherapists.
- Developed muscle and fat segmentation (Dice 0.94 on the internal benchmark) and deployed a distilled MobileSAM-based model under 35 MB.
- Coordinated a six-person data team on a dataset of more than 13,000 images.
Python, PyTorch, OpenCV, MediaPipe, ViTPose+, MobileSAM, ONNX Runtime, TensorRT, CVAT, Labelbox, FastAPI
Computer Vision Engineer
Feb 2022 to Dec 2023SoftwareMotion, Suzhou, China. Remote, project-based.
- Built risk-aware multi-object tracking with collision-course estimation for ADAS warnings.
- Built real-time driver monitoring: gaze, blink rate, head pose and seatbelt use.
- Delivered a drowsiness model that runs under 15 ms on Jetson Orin Nano.
- Prototyped bird's-eye-view fusion from several cameras with a lightweight LSS-style lifter.
Python, C++, PyTorch, OpenCV, YOLO, MediaPipe, DeepStream, TensorRT, ONNX Runtime, Jetson Orin
Computer Vision Researcher
Sep 2020 to Dec 2021G42 Healthcare, Dubai, UAE. Remote, part-time.
- Built weakly supervised 3D detection and segmentation for MRI and CT volumes using pseudo-labels.
- Developed a CNN-Transformer model for tooth detection in panoramic dental X-rays.
- Deployed skin-lesion triage on Jetson Nano; INT8 inference ran up to four times faster than FP32.
- Reduced compute for brain-tumour detection by about 35% with entropy-guided slice selection.
Python, PyTorch, MONAI, U-Net, CNN-Transformers, ONNX Runtime, INT8 quantization, Jetson Nano
M.Sc. Computer Engineering, Artificial Intelligence
Sep 2018 to Feb 2022Amirkabir University of Technology (Tehran Polytechnic). GPA 3.86 of 4.
Thesis: Few-shot Cascade R-CNN for Human Forensic Age Estimation.
AI Computer Vision Developer
Feb 2016 to Mar 2019AMA-CTRL, Shiraz, Iran. On-site.
- Developed a real-time odometer-reading system for industrial meter monitoring.
- Built helmet detection with face recognition for site access, and pose checks for harness use and restricted zones.
- Delivered real-time monitoring of workers and vehicles with tracking and heatmaps.
- Automated blood-test report digitization with Tesseract OCR, reducing manual data entry by 70%.
Python, OpenCV, YOLO, TensorFlow, PyTorch, Tesseract OCR, Raspberry Pi, Jetson
All projects
Client work, research and open source in one list. Pick a field, or type a word such as thermal, Jetson or RAG.
Labs, campuses and conferences
Select a photo to enlarge it.
Tools
- Computer vision
- Object detection, Multi-object tracking, Instance and semantic segmentation, Pose estimation, OCR, 3D vision, Small-object detection, OpenCV, MMDetection, MONAI, MediaPipe
- Deep learning
- PyTorch, TensorFlow, Keras, scikit-learn, Transformers, U-Net, R-CNN family, YOLO, Few-shot learning, Weakly supervised learning, Metric learning
- Foundation models and LLMs
- CLIP, DINOv2, SigLIP, Grounding DINO, SAM 2, SAM 3, Hugging Face Transformers, LangChain, LangGraph, LlamaIndex, RAG, PEFT and LoRA
- Deployment
- ONNX Runtime, TensorRT, INT8 quantization, NVIDIA DeepStream, TensorFlow Lite, Docker, FastAPI, MLflow, Weights & Biases, Comet, Neptune
- Geospatial and data
- GDAL, Rasterio, GeoPandas, TorchGeo, SAHI, Sentinel Hub, Prithvi, Clay, CVAT, Labelbox, Roboflow
- Languages and platforms
- Python, C++, TypeScript, SQL, Linux, AWS, Azure, NVIDIA Jetson, Raspberry Pi
Education and certificates
Each certificate links to its verification page.
Education
- M.Sc. Computer Engineering, Artificial IntelligenceAmirkabir University of Technology (Tehran Polytechnic), 2018 to 2022. GPA 3.86 of 4. Thesis: Few-shot Cascade R-CNN for Human Forensic Age Estimation.
- B.Sc. Computer EngineeringFasa University, 2013 to 2017. Thesis: 3D Human Body Reconstruction from Clothed Images.
Languages
- English, fluentSmallTalk2 speaking test certificate
- French, DELF B1
- Persian, native
Get in touch
Open to senior computer vision and machine learning roles, research collaborations and PhD opportunities.
amirsoltani.os@gmail.com