Oney Erge

Applied AI and Physics, Research and Engineering

PhD, UT Austin
32+ papers
8 patents
800+ citations
Houston, Texas

I enjoy the intersection of AI, physics, and engineering. I've worked across agents, robotics, simulation, machine learning, and physical modeling, and I'm especially drawn to emerging problems where scientific ideas can lead to practical results.

Now

These are some of the projects I work on in public, from local agents and full-precision inference on small GPUs to simulation environments where agents learn from physical outcomes. I hope they are useful to you in one way or another.

Writing

Field note / 01Model memory map
Qwen3-14B · BF16 weights
29.5 GB
Measured against an 8 GB laptop GPU
Oney ErgeApplied AI + systems
Running large language models on small GPUs · Part 1September 13, 202612 min read
Memory requirements of language model inference on a consumer GPU
What occupies GPU memory while a language model runs, calculated for Qwen3-14B on an 8 GB laptop GPU: system use, number formats, model parts, and the KV cache.
- GPU memory
- LLM inference
- BF16
Read the essay

Browse all writing

Projects

Selected work

Policy protected
Agent systems
YBM
A policy-protected local agent that keeps working toward the task.
Talk to YBM from the web, Telegram, or WhatsApp. Tools pass through policy, approval, verification, and receipts. When a needed capability is missing, it can write and run the code required to continue.
Local-first · Self-extending workflows · Auditable tool use
- Policy-gated tools
- Self-extending workflows
- Local-first
Source
29.5 GB modelon8 GB GPU
Lossless weight streaming
Inference systems
Afterimage
Run large, full-precision models on smaller GPUs.
Afterimage is a research toolbox for full-precision inference beyond VRAM. It ran a measured 29.5 GB Qwen3-14B BF16 model on an 8 GB GPU and compares compression, streaming, and speculative methods against named baselines. Its fastest tested mode reached 3.15x AirLLM and 1.56x HF Accelerate.
29.5 GB model · 8 GB GPU · Lossless weights
- Lossless compression
- CUDA memory streaming
- Inference research
Source
Choose · deploy · measure
Local AI
LocalDeploy
A complete local model deployment and benchmarking suite.
LocalDeploy inspects your hardware, estimates which models fit, lets you find and pull them from one UI, manages compatible runtimes, and measures the speed, quality, and memory use you actually get.
Hardware fit · Runtime control · Repeatable benchmarks
- Hardware-aware deployment
- Model benchmarking
- Runtime orchestration
Source
CITY_PULSE / SCORE 100%
Agent evaluation
AgentGymnasium
Bring AI agents and physics into the same experiment.
AgentGymnasium gives an agent a simulated world, explicit tools, and physical challenges, then measures what it builds, what happens, and how its next attempt changes.
24 explicit tools · Explainable scores · Replayable trials
- Agent evaluation
- Physics simulation
- Reproducible trials
Source
RUN / OBSERVE / DIAGNOSE
Robotics
Creature Lab
Design, evolve, and test robots in a reproducible physics lab.
Creature Lab lets you modify humanoids, quadrupeds, and other robot bodies, evolve morphology and control, diagnose failures, and package the complete experiment for another person to reproduce.
Humanoids and quadrupeds · Evolution · PyBullet
- Robot morphology
- Evolutionary search
- PyBullet
Source
CONFIG / TRAIN / INFER
Computer vision
SegCraft
Train, compare, and run semantic segmentation from one platform.
SegCraft includes ready-to-use presets and model backends for training, evaluation, and comparison, then carries the selected setup into image, uploaded video, and YouTube inference workflows.
Model presets · Comparative evaluation · Video workflows
- Semantic segmentation
- Model comparison
- Video inference
Source

Past work

2025 to now
WorkOxy
AI Lead, Language Modeling Engineering
I lead NLP and language-model engineering in Oxy's Applied AI Center of Excellence.
2023 to 2025
WorkDeloitte
AI Lead
I was responsible for 19 models from problem definition through deployment, spanning machine learning, NLP, and optimization.
2015 to 2025
Editorial serviceJGSE
Associate Editor
I served for more than 10 years as an Associate Editor, overseeing peer review and editorial decisions for more than 250 scientific manuscripts.
2022 to 2023
WorkDow
Senior Research Specialist
I published 4 internal technical papers on signal processing, data-driven modeling, and decision-making. I also led the assessment and integration of Boston Dynamics Spot and other quadruped robotic platforms.
2021 to 2022
WorkRobotics Startup
Data Scientist / Robotics Software Engineer
I developed a custom SLAM and sensor-fusion method for an omnidirectional mobile robot, using motor effort and onboard sensing to estimate and correct motion drift for better pose estimation.
2019 to 2021
Research and educationThe University of Texas at Austin
PhD, Dynamic Systems and Control · Portfolio in Applied Statistical Modeling
My doctoral research focused on hybrid modeling: combining AI and physics to improve prediction and control in physical systems. Doctoral dissertation SciPy paper
2013 to 2019
WorkSLB
Data Scientist / Software Engineer
I worked across the full software and data-science lifecycle of ABBL, from pattern-recognition models through product development and commercialization. The automated directional-drilling adviser was tested across more than 250 wells. ABBL service
2011 to 2013
Research and educationThe University of Tulsa
MSc, Petroleum Engineering
My MSc research combined experiments and numerical methods for non-Newtonian flow and freely rotating drillstrings. I later collaborated on work pairing a deep-learning model with genetic-algorithm optimization. Request master's thesis Genetic-algorithm paper
2006 to 2010
EducationMiddle East Technical University
BSc, Petroleum and Natural Gas Engineering
I built my engineering foundation in transport phenomena, numerical methods, and subsurface systems.