Nick Popelnukh

Projects

Work at Kids Jump Tech, shipped to production hardware.

Diagram: RGB and depth frames fused into a 3D body skeleton at about 70 FPS

Real-time 3D body tracking with depth fusion PyTorch, ROMP, Intel RealSense, AWS SageMaker

Extended ROMP with a depth branch that fuses RealSense depth and RGB features. Runs at ~70 FPS and shipped to 3,000+ hardware units on low-profile RTX 5050 GPUs.

Diagram: progressively darkened frames and a 29% lower MPJPE bar for the RGB-D model versus RGB-only

Low-light 3D pose estimation from synthetic darkening SMPL fitting, RGB-D, MPJPE

Fine-tuned on ~40K self-generated, synthetically darkened RGB-D frames with SMPL fitting. Cut low-light pose error by ~29% (MPJPE) relative to an RGB-only baseline.

Hand-joint skeleton with 10 mm versus 15 mm low-light error bars

Depth-fused hand tracking for custom gestures RTMPose, knowledge distillation, RGB-D

Distilled a student model from RTMPose that fuses depth with darkened RGB. In low light it reaches 10 mm root-relative joint error versus 15 mm for RGB-only (~33% lower) at ~30 ms. Because it tracks joints instead of classifying fixed gestures, users can define unlimited gestures.

Bar chart: sensor pipeline latency reduced from 140 ms to 10 to 50 ms

Inference and pipeline optimization ONNX, TensorRT, FP16, C++, C#, Python

Sped up model inference 2–5x with FP16 TensorRT exports and no visible accuracy regression in visual QA. Profiled and optimized the sensor pipeline, cutting latency 64–93% (140 ms down to 10–50 ms).

A human skeleton retargeted onto several different character rigs

Live human-to-character motion capture and LiDAR touch Unity, Python, LiDAR

Built the company's first live motion capture feature, retargeting tracked human motion onto any T-pose rig (20+ rigs). Also built the pipeline that turns LiDAR wall and floor touches into Unity touch events.

Diagram of a sensor, projector and floor plane aligned for projection mapping

In-house sensor and projection calibration Extrinsics, floor-plane alignment, projection mapping

Replaced a deprecated third-party tool with software for sensor extrinsics alignment, floor-plane alignment, and display/projection mapping. It is used in customer setup and saves an estimated ~$600K/year in licensing.

Pipeline diagram: documents to vector index to LLM to marketplace listing

LLM-powered marketplace automation RAG, OpenSearch, PostgreSQL, Next.js

At Classeasily, built a RAG pipeline and LLM automation that keeps marketplace listings current and analyzes customer reviews, running in production. Also cut search latency 92–94% (~5 s to ~0.3–0.4 s) by moving from SQL joins to OpenSearch.

Experience

Nov 2023 – Oct 2026

Machine Learning Engineer, Kids Jump Tech
Interactive floor, wall, and table systems for schools and entertainment venues. Miami, FL (hybrid)

Jan 2025 – Aug 2026

Founding Lead Engineer (part-time), Classeasily
Marketplace platform reaching 20,000+ people, built with a team of 6. Toronto, ON (remote)

Aug 2025 – Oct 2026

Software Engineer (part-time), Confident Care of Florida
Event-driven AWS Lambda automation and company website. Jacksonville, FL (hybrid)

Skills

Languages: Python, C++, C#, SQL, TypeScript
ML and real-time: PyTorch, TensorFlow/Keras, scikit-learn, OpenCV, CUDA, ONNX, TensorRT, FP16 quantization, ROMP, SMPL, RTMPose, YOLO, MediaPipe, Detectron, Intel RealSense, LiDAR
MLOps: AWS SageMaker, NVIDIA Jetson, MLflow, Weights & Biases, Docker, Kubernetes, Terraform, GitHub Actions
LLMs: Hugging Face Transformers, LoRA fine-tuning, RAG and vector databases, OpenAI and Anthropic APIs
Backend and web: Node.js, REST APIs, PostgreSQL, OpenSearch, AWS (Lambda, EKS, S3), React, Next.js

Projects

Experience

Education

Skills