Argo Saakyan — Lead Computer Vision Engineer

Summary

Building real-time perception for Physical AI at Agnify (Austin, TX) — a GPU-resident TensorRT pipeline serving 14 live CCTV cameras (~400M frames/month), benchmarked to 350 fps on a single GPU.
Tech lead and core owner of fraud detection at Veryfi (YC17), built and deployed 13 vision models, each processing 5–12 million images per month.
Main author of D-FINE-seg, an open-source object detection and instance segmentation framework with paper + GitHub — now the core of Agnify's perception stack.
Strong in deployment and edge inference: TensorRT, OpenVINO, ONNX, CoreML, Triton Inference Server, Linux, hardware setup, physical prototyping.
Led ML engineers and coordinated data/annotation workflows.
Inspired by robotics, self-driving, drones, space, agriculture, hardware projects, edge devices.
Speaking at conferences, writing articles about CV/DL, judging international awards.
Hands-on with hardware and prototyping: Linux workstations, microcontrollers, single-board computers, sensors, woodworking, and metalworking.

                        6
                        Years Experience
                    

                        20+
                        Models Shipped
                    

                        5M+
                        Images / Month
                    

                        400M
                        Frames / Month
                    

Experience

Senior Computer Vision Engineer

Agnify — Austin, Texas

Apr 2026 — Present

Physical AI startup giving factories real-time visual understanding of their production lines.

Built the core perception stack as author of D-FINE-seg — object detection + instance segmentation.
Engineered a fully GPU-resident TensorRT pipeline (NVDEC → preprocess → inference → postprocess → NVENC) serving 14 live CCTV cameras (~400M frames/month) — benchmarked to 350 fps on a single GPU, capped at 10 fps/camera to free compute for downstream tasks.
Trained, optimized, and deployed specialized models end-to-end, including data mining.
Built features: object tracking, counting, pose estimation, and a SAM3 auto-labeling solution.
Achieved 99.8% accuracy on a bottle-counting task (D-FINE-seg → tracker → counter).

PyTorchTensorRTSAM3ByteTrackQwen (VLM)NVDEC/NVENCDockerLinux

Previously

Lead Computer Vision Engineer

Veryfi (YC17) — Silicon Valley

Jun 2023 — May 2026

Tech Lead of the Fraud Detection suite; built the decision engine fusing vision + text model signals into a final fraud score.

Led a team of fraud-detection ML engineers and coordinated delivery across teams.
Built and deployed 13 production vision models (classification, detection, segmentation), each processing 5–12M images/month.
Designed custom conv/transformer architectures beating prior models on both F1 and latency (raised one from 93% to 99%, F1 0.98).
Shipped 700KB on-device models (limit 1.5MB) at 7fps; ONNX / TorchScript / TensorRT / TFLite backends.

PyTorchTensorRTTorchScriptTFLiteONNX

CV Engineer & Adviser Retinal pathology segmentation, real-time billiards analysis (60fps), factory defect detection

CV Researcher — Diagnocat 3D dental imaging, medical AI deployed across multiple countries

CV Engineer — DisArm Gun detection, real-time video surveillance, edge deployment

DL Engineer — BigDataMSU Infrastructure monitoring, industrial anomaly detection, hardware prototyping

Data Scientist / Analyst Healthcare analytics, insurance, forecasting & automation

Founder — Earthome IoT smart home systems, hardware prototyping, 10+ deployments

Open Source & Projects

Featured

D-FINE-seg

Object Detection & Instance Segmentation Framework

End-to-end framework for training, exporting, and running inference with state-of-the-art object detection and instance segmentation models. Published with an accompanying research paper, and now the core of Agnify's production perception stack — 14 live cameras, 99.8% counting accuracy.

GitHub Paper

PyTorchObject DetectionSegmentationONNXTensorRT

Triton Server Pipeline

Open-source inference pipeline built on NVIDIA Triton Inference Server for scalable, optimized model serving in production environments.

GitHub

TritonTensorRTDockergRPC

PyTorch Training Pipeline

Reusable, modular training pipeline for computer vision tasks — designed for rapid experimentation and clean reproducibility.

GitHub

PyTorchHydraAlbumentationstimm

Hardware & Prototyping

Physical prototypes with microcontrollers, single-board computers, and sensors (lidar, radar, ultrasonic). Linux workstation builds. Woodworking & metalworking.

Nvidia JetsonRaspberry PiArduinoSensorsLinux

Technical Skills

Deep Learning & CV

PyTorch TorchVision timm / HuggingFace OpenCV Albumentations ByteTrack SAM 3 VLMs (Qwen) Lightning AI NumPy Scikit-Learn

Optimization & Deployment

TensorRT OpenVINO ONNX CoreML TensorFlow Lite TorchScript Triton Inference Server NVDEC / NVENC

Infrastructure & Tools

Linux Docker Git Flask FastAPI Hydra Python

Hardware & Prototyping

Linux Workstations Nvidia Jetson Microcontrollers Single-Board Computers Soldering Lidar / Radar / Ultrasonic Woodworking Metalworking

Domains

Object Detection Instance Segmentation 3D Segmentation Classification Object Tracking Pose Estimation Counting Anomaly Detection Medical Imaging Edge Inference

Speaking, Writing & Awards

Conferences

DevDays Europe Voxel51 Warsaw IT Days

Published in

ArXiv TowardsDataScience TowardsAI DevGenius

Awards

CODiE Award Judge Hackathon Judge Digital Breakthrough Winner

Let's Connect

Interested in working together, have a project in mind, or just want to chat about computer vision, robotics, or edge AI? I'd love to hear from you.

LinkedIn [email protected] GitHub