Here are the code, test results, and project notes behind the work on my homepage. I can share the personal projects in detail; the enterprise summaries explain my contribution without publishing private company records.
01PERSONAL PROJECT / PUBLIC CODE
ChangeFront: keeping recommendations tied to their sources
THE PROBLEM
What happens to an AI recommendation when its source changes?
I built ChangeFront to keep track of that connection. It records the source versions behind recommendations and document changes, then invalidates the work if a source changes. An AI client can draft through MCP, but a person has to approve it.
This is a local, single-user demo with synthetic data and simulated reviewer identities. It is not production authentication or an enterprise deployment.
02ENTERPRISE WORK / PRIVATE SUPPORTING RECORDS
Enterprise AI architecture and rollout
These summaries draw on implementation, approval, support, and stakeholder records reviewed on August 28, 2026. The originals are private. I can share my approach here, but the site does not provide independent access to those records.
CONTROLLED AZURE AI FOUNDRY PILOT
Testing changed the plan for expanding access.
I designed and provisioned a cloud AI pilot around the team's research, security, access, and export-control requirements. I also coordinated vendor approvals, built the delivery plan, and tested access from a participant's perspective.
What I found
The participant workflow could reach an external-search route outside the intended data boundary.
What I did
I restricted participant access and held off on expanding the pilot pending a safer design.
Where it stood
A working restricted pilot, approvals for specialized controls and model capacity, and a fully signed cross-functional plan. The records do not establish a scaled production rollout or measured business impact.
ENTERPRISE OPERATIONS AND ENABLEMENT
Helping employees move into supported AI tools.
Rollout and migration
I deployed an internal GPT across the workspace and helped a business team move from an early ChatGPT pilot into the corporate workspace. The stakeholder confirmed the support was successful.
Administration and support
I implemented ChatGPT and Codex safeguards across an administrative scope of 100+ users, published usage guidance, and supported employees. That figure is not a count of active users or a measure of adoption.
Governance
I helped shape leadership-approved usage rules and exceptions, coordinated platform and vendor reviews, and served as the internal point person for a defined governance discovery effort.
Project notes and supporting records
These notes explain my approach and the types of records behind the summaries, without publishing private system details.
I worked on the software, electronics, and mechanical design.
I co-developed the Python gait and navigation code, connected the Arduino/C++ sensor hub to the Raspberry Pi, and built simulation and telemetry tools. I owned the CAD design and custom fabrication.
I also connected Claude Code through MCP so I could describe what I saw the rover doing and use its code and telemetry to help find possible causes. Claude assisted with diagnosis; the rover's deterministic software controlled its movement.
REPORTED TEST RESULTS
31/32successful formal traversals
7terrain categories
40/40reported simulation checks
3COSGC recognitions
This was a small field study. Abrupt changes between terrain types remained a problem, and the report notes that an exact software-version hash was not saved for each trial.
See the rover in action.
These are original project recordings. They show the rover's behavior, but are separate from the formal trial results and the code comparison below.
One sensor reading was enough to trigger a danger response.
A regression test captures the case: the front-center sensor reads 5.83 cm while the other two front sensors read 100 cm. The old code takes the most severe reading, so it reports DANGER. The March 19 fix uses the middle of the three severity values, while keeping the danger response when a majority agree or a sensor reports the blind-zone signal.
There is a tradeoff here. Filtering an isolated reading can also miss a real obstacle seen by only one sensor. The tests check which readings are filtered and which still trigger a danger response.
For this September 14, 2026 comparison, the same 15 assertions were run against the classifier functions from three saved code versions. They cover 13 distinct input frames. The older code misses six of the fix's new expectations; that does not mean its original test suite failed. This checks the functions, not the full rover or a new field run. The fix is recorded under my name in the public commit history.
Download the script above. With Python 3.9 or later, run python3 -B rover-voting-evidence.py --output result.json from its folder. It downloads the exact code versions used here and runs only the selected functions, without loading the rover hardware code. On Windows, use python if python3 is unavailable.
I'm looking for an AI Solutions Lead role where I can work on architecture, governance, and implementation. I'd be happy to walk through these projects and the decisions behind them.