01
Stealth startup
Founder
2026 – present
Unannounced
- An early-stage company with machine learning at its core. More once we launch.
- Happy to talk about it — email me.
Durham, NC
Benjamin
Chauhan
Second-time founder, deep learning researcher by training. My first company taught 200,000 students for free; my second is in stealth. In between: research at PNNL, SLAC, and Duke, all of it shipping into production systems.
This is where I keep what I'm working on and what I've built.
Discipline
Deep learning research & ML systems
Currently
Founder, stealth startup · PNNL · Duke MS '26
Now — 2026
Building something new
An early-stage company with machine learning at its core. More soon.
Ask me about it →Also currently
Deep learning research intern
Pacific Northwest National Lab — map understanding and georeferencing at corpus scale.
M.S. Electrical Engineering
Duke University, ML track — finishing May 2026.
Recent
Feb 2026
1st place, NVIDIA challenge — MIT iQuHACK
2025
NEPATEC 2.5 pipeline shipped — 1.2M+ images
2025
President, Duke IEEE Computer Society
01 — Selected work
Roles across four national labs, a hospital lab, and companies of my own.
From labs to startups
02
Pacific Northwest National Lab
Deep Learning Research Intern
Seattle, WA · Jun 2025 – May 2026
- Built, tested and integrated a map image classification pipeline scaling to 1.2M+ permitting images in NEPATEC 2.5.
- Owned the whole pipeline — annotation guide, labelled data collection, and scale-out to the full NEPATEC corpus.
- Building an automatic georeferencing framework that turns scanned permitting maps into georeferenced GeoJSON layers.
03
Duke General Robotics Lab
Deep Learning Researcher
Durham, NC · Sep 2025 – May 2026
- Built an automated YOLO + VLLM pipeline to detect privacy leakage across large-scale robot learning datasets.
- Surfaced real leaks — passwords, emails, phone numbers — inside the DROID robotics dataset, demonstrating the problem at scale.
04
SLAC National Lab
Researcher — E-log Chatbot
Stanford, CA · Jul – Sep 2024
- Built a containerized RAG chatbot for the FACET-II lab on Open WebUI, Langfuse and RagFlow, letting operators ask complex questions about lab operations; contributed upstream to Open WebUI.
- The system is being extended across all of SLAC — the entire e-log of every accelerator run.
05
- Founded and led an AI education startup that gave 200,000+ high school students free personalized AP tutoring, partnering with creators like Heimler's History and answering 3M+ student questions.
- Designed and built a lightweight, scalable RAG agent API on FastAPI and Weaviate holding 98.82% uptime.
06
Self-Driving Golf Cart
Founder, ML specialist — ongoing initiative
Hartsville, SC · 2022 – present
- Designed and built a modular fully autonomous golf cart system, turning retired carts into safe, eco-friendly transport.
- Fused stereoscopic cameras, LIDAR, YOLOv7 and GPS into a driving stack that gave the cart real autonomy.
- Secured $35,000 in grants from Google and the SC Department of Education — and praise from Boston Dynamics.
07
UofSC Research Computing
Deep Learning Research Intern
Jun – Jul 2024
- Evaluated DeepSolo and GoMatching for scene-text detection on USC's historical film archive.
- Raised DeepSolo's grayscale performance by 61% by re-training and fine-tuning on grayscale imagery.
08
Duke Hospital Brain Science Engineering Lab
Research Engineer
Durham, NC · 2022 – 2024
- Designed and tested an open-source computer vision platform for transcranial magnetic stimulation — 100× cheaper than current options.
09
UofSC Advanced Research Computing Lab
Junior Researcher
Columbia, SC · 2021
- Revealed accuracy differences between deep learning frameworks on image classification — a new research direction for the ARC Lab.
- Authored a paper for the South Carolina Junior Academy of Sciences, taking 3rd place in its category.
02 — Projects
Selected projects.
Quantum · 1st place
Feb 2026
iQuHACK 26
A hybrid quantum/classical solver for the low-autocorrelation binary sequence (LABS) problem. First place in the NVIDIA challenge.
CUDA-Q
Optimization
Architecture research
May 2024
CHopT / CHAD
Hopfield networks integrated into LLMs to carry memory across separate inference instances.
PyTorch
Associative memory
Open source
Oct 2024
SimpleRAG
A minimal repository that stands up a working RAG chatbot in a few lines of code.
Python
Retrieval
Infrastructure
May 2024
Audio2Blog
A serverless Rust microarchitecture on AWS that turns recorded conversations into written blog posts.
Rust
AWS Lambda
03 — Research output
Papers, contributions, and coursework.
Paper · 2021
Accuracy differences across deep learning frameworks in image classification
South Carolina Junior Academy of Sciences — 3rd place in category. Opened a new research direction for the UofSC ARC Lab.
SCJAS
Paper · 2024
CHopT (Continual Hopfield Transformers): Preserving Memory Across Inferences ↓
With Peter Liu. Augments Llama 3.2 3B with a learnable, gated Hopfield memory layer to retain knowledge across inference boundaries. Duke ECE 661.
Duke ECE
Report · 2024
Investigating Architectural Placement of Hopfield Memory Layers in CHopT for Improved Fluency ↓
Solo follow-up testing earlier memory-layer integration in the Llama 3.2 3B stack — found it destabilizes fluency under constrained LoRA fine-tuning.
Duke
Project · 2026
Benchmarking Sparse Attention for Efficient Image and Video Synthesis ↓
Two-phase study benchmarking sparse attention kernels (Block Sparsity, ADSA, Re-ttention, PISA, SparseVideoGen2, VSA) against dense attention in Diffusion Transformers, on A6000 and H100 GPUs.
ECE 590
In progress · 2026
Privacy leakage in large-scale robot learning datasets
Automated detection of passwords, emails and phone numbers in DROID, using a YOLO + VLLM pipeline. Duke General Robotics Lab.
Duke GRL
Open source · 2024
Contributions to Open WebUI
Upstream work from building SLAC's FACET-II e-log chatbot on Open WebUI, Langfuse and RagFlow.
SLAC
Coursework
ML Algorithms · ML Systems · Deep Learning · Transformer Architecture · Robot Learning · AI Security
B.S. Computer Science '25 and M.S. Electrical Engineering, Machine Learning track '26 — Duke University, GPA 3.925.
Duke
04 — Leadership & teaching
Leadership and teaching.
President · Jan 2025 – present
Duke IEEE Computer Society
- Ran 20+ events connecting Duke students to industry across technology sectors.
- Oversaw conference travel for 25+ students — Supercomputing 25, IEEE SoutheastCon 25, SatShow 26.
Vice President · Jan 2023 – present
Duke Robotics Mentorship
- Managed the Durham school system relationship, transportation and a $4,000 budget.
- Coordinated lesson plans across 4 middle schools for 80+ students; secured $1,500 in additional grants.
Student teacher · 2020 – 2022
SCGSSM Spark!
- Planned and taught interactive STEM lessons to 60+ middle schoolers across South Carolina.
- Directed two teams of student teachers and two live lessons.
05 — Beyond the lab
Other things I do.
Teaching
I like teaching this stuff
I taught middle schoolers in South Carolina, mentor robotics students in Durham, and built an AP tutor that ended up helping about 200,000 kids. Explaining something is usually how I find out whether I actually understand it.
Hands on hardware
I like building physical things
A friend and I made a golf cart drive itself in a barn in South Carolina. At Duke I helped build a camera setup for a brain-stimulation lab that cost about a hundred times less than the commercial one. Software is great, but I like it when there is something to point at.
Languages
Polish and Chinese
Polish is fluent — it is what I grew up speaking at home. Chinese I can hold a conversation in, and I am still working on it.
Things I keep wondering about
- What else is hiding in the datasets everyone trains on.
- Why so much public data — old maps, lab logs, film archives — is still impossible to search.
- Whether you can trust a benchmark before you have looked at the data yourself.
Drop an image — teaching, workshop, or travel
06 — About
Hello.
Most of my work has started with an archive nobody could search: a million scanned permitting maps, decades of accelerator logs, a robotics dataset nobody had audited. The modeling is rarely the hard part — the hard part is the annotation guide, the container, the eval that tells you the truth.
I finish a masters in Electrical Engineering at Duke in May 2026, on the machine learning track, after a CS degree there. Before that I built an AI tutoring startup out of a dorm room and an autonomous golf cart out of a barn in South Carolina.
Fluent in Polish, conversational in Chinese, happiest with a profiler open.
Based in
Durham, NC
Awards
Dean's List · FIRST SC · ACCESS CCEP
07 — Stack
Tools I work in.
Modeling
PyTorch
TensorFlow
Detectron2
YOLO
VLLM
Languages
Python
Rust
SQL
Systems & ops
Docker
Compose
Apptainer
K8s
MLOps
HPC DevOps
Cloud & services
FastAPI
AWS Lambda
EC2
Bedrock
S3
Weaviate
Drop an image — field work, hardware, or the cart
Drop an image — desk, terminal, or whiteboard
08 — Contact
Ask me what I'm building.
Elsewhere
Full résumé
Download PDF ↓