Advanced Metal Research
Learning and autonomy engineer on robotic welding (Hawthorne, July 2026 to present, through Mabie Industries). I work across the co-founders to turn competing priorities into one engineering plan covering robot motion and welding, run data, robot CI/CD and deployment, and weld procedure integration, and I lead three to five of the team’s interns.
I delivered the shared run-data platform for the laser welder and the robot cells, from the robot controller to queryable evidence: typed robot-state rows, content-addressed capture, and a custody chain with explicit failure cases. Weld books, materials, joints, process windows, trajectories, and evidence are typed models, and each run is bound to source revision, binaries, robot model, calibration, and collision assets, so a result can be re-derived rather than trusted.
I wrote the team’s verification ladder, from hooks and lint through simulator, hardware, and weld-sample proofs to runtime monitors, and tied merges and releases to it. I also built the team’s agent software factory: Codex and Claude agents under one set of responsibilities, cross-agent review, a written verdict on every artifact, and simulated evidence never reported as physical.
On the motion side, torch orientation and robot/positioner coordination are optimization variables under independent kinematic, collision, calibration, and process checks. I designed the staged path from simulation to supervised process-off trials, and built Steam Deck operator controls and runtime integration on the way to operator-confirmed jogging on the assembled robot. Full trajectory execution and welding remain separate validation work.
Applied Medical
From sixteen to twenty-one, intern and then full-time engineer in medical-device manufacturing. I designed, built, and qualified dozens of end-of-arm tools, fixtures, and other tooling for polymer injection molding and its automation, and they still run in production daily.
One metal-injection insert-molding problem was losing about $100k a month to pin misalignment on a new suturing device. I traced it to a magnetic pen polarizing the pins, then built a 3D-printed orientation grid with fishing-line sorting and machined stainless suction-pen tips, which ended the variance in production.
T-UEBA
Principal investigator on a $250k Army SBIR at RAM Laboratories in tactical user and entity behavior analytics, producing enforcement rules for zero-trust policy points. I wrote the winning proposal, led one full-time engineer, and owned the graph schema and ontology, the synthetic-data generator for operational phase shifts, the data pipeline, the model, and the evaluation.
The model was a heterogeneous temporal graph network with calibrated graph attention, conformal thresholds, and sparse-autoencoder inspection. Results: 95% accuracy, 0.887 F1, under 3% false positives, 15.2 ms per graph, and a 15 MB model, 400× smaller and 7.9× faster than the baseline. In evaluation it generated candidate enforcement rules for policy decision and enforcement points.
CAD from drawings
A work trial with Augrade on converting DXF drawings into editable CAD: vector-native tokenization, provenance-preserving merges, and hard validators outside the learned stack. A public standard-library Python workbench recovered 2,237 wall, column, and curtain-wall polygons from about 67,000 primitives across 111 layers, with parameter sweeps and SVG review artifacts.
Automated code repair
DL-Patcher, an MDA SBIR on automated C/C++ vulnerability repair: I joined in Phase II and moved the models from seq2seq to RoBERTa, where repair started to work, with synthetic bug insertion, mined natural patches, static-checker feedback, patch verification, and a human-in-the-loop VS Code workflow. I’m a named co-inventor on the training method, patent pending (U.S. application 18/375,839).
I was also a core contributor to VULCAN, RAM’s DARPA AI Cyber Challenge entry: a code-model harness where static analysis, decompilation, fuzzing, and runtime verification supply the accept or reject signal on generated patches.
Production systems
Sole engineer at Drone Institute, through Mabie Industries: I built the company’s software from scratch, six production services (lead marketplace, partner upload portal, drone orders, operations dashboard, thermal-imagery lead generation, pilot dashboard) and the CI/CD, infrastructure, and tests behind them. Serverless AWS with per-service dev and production deploys, running at about $30 a month.