High school student working on aerospace simulation, autonomy, and software. I like projects where the result can be checked against data, a reference implementation, or a frozen test set.
I built AegisLand to test a narrow question: if a landing camera is confidently wrong, can an independent estimate expose the error before touchdown?
An early synthetic experiment looked promising: unsafe touchdowns fell from 43% to 1% with selective intervention. Later camera-based testing was less clean. On the Phase 10R frozen holdout, average error improved, but the system missed 20% of truth-visible frames, failed both p95 tail-error targets, and under-covered its stated 95% uncertainty interval (84.3% lateral / 79.7% altitude). I kept that result frozen instead of tuning on the test set.
Repo: uav-safety-research
I have been contributing to Elodin's Voyager example. My first merged contribution added position/velocity error telemetry against SPICE reference data. I then worked on a reconstructed Voyager 1 Jupiter validation case and on keeping the assumptions and limitations visible in the example instead of hiding them behind a large simulation stack.
Merged example contribution: elodin-sys/elodin#769
I am also working through public spacecraft dynamics and controls results one at a time. For a Lockheed Martin HWO / disturbance-free-payload paper, I used a one-degree-of-freedom spring-mass model to check whether the reported modal-frequency shift is consistent with a 70% effective stiffness cancellation. The simple model gives the expected roughly 1.8x frequency ratio. It does not reproduce the full 6-DOF controller, hardware, or flight design.
That kind of boundary matters to me: if I did not model it, I do not want the project page to imply that I did.
AbstainBench — a small browser benchmark for a simple question: when should a language model answer, and when should it say it does not know? It records correct answers, hallucinations, and false abstentions across a fixed set of prompts. Repo
TennisRank — a private team ranking tool with spreadsheet import, singles/doubles rankings, and Supabase persistence. Repo
ECHO / FIELD — a browser-based visual instrument where microphone input and motion alter a live generative field. Repo
- Seagulls / OpenStage — AI engineering intern. Worked on model routing, memory boundaries, guardrails, tracing, QA, and production-path checks. Recent work included looking for inference paths that bypassed the intended routing layer and testing user-facing failure cases.
- Open-source aerospace work. Contributing to Elodin and exploring public NASA/Aviary work through reproducible examples and validation tasks.
- Student research and engineering. Independent simulation work plus school and community engineering projects.
River Islands High School · Delta College coursework
SkillsUSA regional champion in 3D Visualization & Animation; advanced to the California state competition. Also placed 2nd in a Python game jam and 5th of 25+ teams at an AI Collective hackathon.
I use Python and JavaScript most often. For simulation/research work I also use NumPy, Matplotlib, Git, and the surrounding domain tools as needed.
I would rather show a small result I can explain than a large system description I cannot defend.