Hi, I'm Ron.

I build spatial AI systems.

Ron Kibel

Experience

Aurora

Software Engineer - Vehicle Platforms

August 2026 → Present

Building and validating the hardware-software interface for autonomous trucks.

Google

Student Researcher - Google Beam

May → September 2025

Engineered camera ML tooling for a zero-copy, hardware-accelerated graphics pipeline.

ServiceNow

SWE Intern - AI Search

May → September 2024

Built multimodal retrieval for text-to-image and image-to-image workflows.

SWE Intern - Developer Productivity

May 2022 → September 2023

Created and deployed static code scan automation for all engineering orgs.

Projects

POINTER workflow from object capture through persistent augmented reality placement

POINTER: A Real-Time Framework for Dynamic, Object-Centered Augmented Reality

Object-centered AR framework that turns everyday objects into persistent digital canvases and tangible controllers. iOS-to-cloud pipeline with 720p + 30 FPS end-to-end flow, composed of WebRTC streaming, semantic extraction, open-vocabulary segmentation, and model-free tracking.

Egocentric cooking scene with gaze, depth, saliency, and object tracking outputs

Temporally-Aware Scene Simplification for Bionic Vision

Prosthetic-vision perception pipeline that preserves hands and task-relevant objects while suppressing background clutter in egocentric daily living. Benchmarked with depth, saliency, segmentation, and hand-prior variants and trained with supervised mask quality and unsupervised temporal stability.

Reconstructed neighborhood point cloud with the simulated drone trajectory

Semantic 3D Reconstruction from Multi-Robot Systems

Graph-based multi-robot framework for depth prediction and semantic segmentation. Tested with 16 drones and 22K AirSim RGBD images + noise, fusing features across drones and reducing training time by 19% per epoch.

Energy efficiency comparisons for ResNet18, LSTM, LeNet, and GRU architectures

AheadOfItsTime: AOT vs JIT Compilation

Benchmark study comparing ahead-of-time and just-in-time compilation across deep-learning runtimes. Performance swings from 2× faster to 50× slower: we derive an architecture and batch aware deployment strategy.

A sequence of video frames flowing into a vision-language model judge that produces a verdict and reward

Mitigating Reward Hacking

Reinforcement-learning study that induced reward hacking through intentional reward misspecification, then used frame-sequence VLM judges to detect and regularize exploitative policies. Detected 100% of hacky MuJoCo trajectories while exposing convergence limits.

Skills

Languages

PythonC++JavaJavaScriptHTML/CSSSQL

Computer Vision / ML

PyTorchTensorFlow/KerasOpenCVCLIPCUDAONNX

Graphics / Systems

OpenGLVA-APIDRMReal-time pipelinesHardware acceleration

Infrastructure / Tools

DockerGCPGitDVCMavenJenkins CI/CD