Robotics · AI/ML · Software
I build and audit intelligent systems, from robot data pipelines and fault injection rigs to reproducible machine learning research.
I am an AI Robot Operator at Physical Intelligence and Founder & Executive Director of House of Seva. I hold an M.S. in Software Engineering and a B.S. in Computer Science.
Open to full time robotics software, AI/ML, and software engineering opportunities.
Current focus
Reproducible work with the claim, evidence, controls, and limitations kept together.
Tests whether apparent state counts from PCA, TICA, and VAE analyses partly reflect the number of saved frames. Controlled synthetic landscapes separate sample size effects from exploration deficits, with validation on two independent 100 ns alanine dipeptide trajectories. Submitted to JCTC on September 15, 2026; the ChemRxiv preprint is under review.
Reproduces a published five model ordering under one pinned RTX 4070 evaluation stack. Four models retained their relative order, while StarCoder2 3B reversed with Qwen2.5 Coder 0.5B, reducing Kendall’s tau b to 0.8. Version 1.0 is a technical report and has not been submitted to a venue.
Finds that INT8 is smaller but not intrinsically faster or accurate. Per channel INT8 accelerated one thread XNNPACK inference by 1.27 times, yet ran 2.24 times slower without the default delegate. Version 1.2.0 is archived on Zenodo and has not been submitted for peer review.
A denominator aware study across ResNet 18, VGG16, and MobileNetV2 with three training seeds, full test evaluation, sensitivity analyses, and prediction level artifacts. Submitted to Transactions on Machine Learning Research through OpenReview.
Finds that every traceable tumor test image belongs to a patient also present in training and that 216 test images are pixel identical to training images. The release provides recovered patient identifiers and patient disjoint folds. The manuscript has not been submitted to a journal.
A controlled study of 660 policy fits across five failure modes, two policy families, ten seeds, and transition matched controls. It shows why open loop prediction error can misrepresent closed loop policy damage. Submitted to the CoRL 2026 Learning from Corrections and Interventions workshop.
Audits SAL and TED on robomimic and finds that the metrics primarily rank operator skill rather than task success. Venue specific versions are under review at CoRL WEBP, NeurIPS RoboPAD, and CoRL Oops, I Erred.
Credentials
Selected technical and professional credentials.
Accelerated Computing in Modern CUDA C++September 2026
HackerRank Software Engineer CertificateMarch 2026
Cisco Network Support and SecuritySeptember 2025
IBM Artificial Intelligence FundamentalsAugust 2025
Stanford Fundamentals of AI and Machine Learning in HealthcareJune 2025
Google Foundations of Project ManagementSeptember 2024
Writing
A three node CAN bus built to induce and diagnose failures that software logs alone cannot explain.
Read on Medium →
Rebuilding the core idea behind GELLO style robot demonstration collection with direct joint space control.
Read on Medium →Timeline
Physical Intelligence · May 2026 to present
House of Seva · June 2026 to present
Handshake AI · April to June 2026
Kaiser Permanente · October 2023 to June 2024
Grand Canyon University · 2024 to 2025
University of Silicon Valley · 2021 to 2023
Beyond work
I read philosophy, theology, science, and the classics. Add me on Goodreads to see what I am reading, and follow my writing on Medium.
I enjoy building custom keyboards and plan to release projects and articles about them whenever possible. Stay tuned.
I design, print, and refine physical pieces through Castline Studio, combining digital fabrication with hands on experimentation.
Build, break, measure
Robotics infrastructure, fault characterization, ML tooling, and production software. More on my GitHub.
Kubernetes · Redis Streams · PostgreSQL
A multi stage pipeline designed for deliberate pod and network failures, with durable acknowledgements, replay safe writes, health probes, observability, and a chaos harness.
GitHub →
URSim · RTDE · Fault injection
Simulator backed control and synchronized telemetry with reproducible tests for connection failures, stale handles, hangs, and native crashes.
GitHub →
SwiftUI · AVAudioEngine · DSP
A dependency free tinnitus habituation app with real time generated sound enrichment, personalized notch filtering, guided exercises, and progress tracking.
GitHub →
Python · Streamlit · Diagnostics
Robot fleet log analysis with fault classification, recurrence detection, anomaly detection, rule based root cause analysis, and a live health dashboard.
GitHub →
Computer vision · Labelbox · COCO
A data operations workflow for robot arm grasp annotations with bounding boxes, keypoints, failure labels, and automated COCO JSON export.
GitHub →
Python · Docker · MLOps
A library and CLI for packaging machine learning models into portable, optimized Docker inference images.
GitHub →
C++ · PyTorch · Self play
An AlphaZero inspired system combining a C++ bitboard engine, PyTorch neural network, self play training, and a PyGame interface.
GitHub →
OpenAI · React · Node.js · Supabase
An AI resume analyzer with job specific feedback, ATS scoring, skill gap analysis, and PDF export.
Live site →
Python · PubMed · NLP
A biomedical literature tool for searching PubMed, rewriting summaries for different audiences, exporting citations, and generating PDFs.
Live site →
Next.js · TypeScript · WebAssembly
An e commerce workflow with client side STL analysis, Etsy and shipping integrations, automated email, and deployment on Vercel.
Live site →