If the camera is sure and wrong, can the vehicle refuse to land?
High school researcher exploring perception reliability, calibrated abstention, and reproducible simulation at the intersection of aerospace, AI, and autonomy. I freeze holdouts — including the ones that fail — so results can be inspected, not just trusted.
🛡️ AegisLand
UAV perception safety research · simulation-only
External perception evidence you can inspect. Preregistered gates, frozen mixed holdouts, and a live research cockpit. Phase 10R improved average error but failed tail, availability, and calibration gates — frozen as-is for Phase 11.
- 43% → 1% unsafe episodes (Phase 6B synthetic, selective confidence)
- 84.2% → 2.4% unsafe rate (V3 redundant perception architecture)
- Phase 10R: mixed/failed result frozen at
e1d566fwithout post-exposure tuning
Case Study · Research Cockpit · GitHub
Note:
safety_acceptance = false. Simulation only. Not flight software.
LLM calibrated abstention benchmark
The same question as AegisLand, for language: when should a model answer, and when should it say it doesn't know? 30-question local-first benchmark with no API keys.
- Measures answer accuracy, abstention accuracy, hallucinations, false refusals
- Offline baseline + optional in-browser Llama 3.2 1B via WebGPU
- JSON and Markdown export for reproducibility
Generative instrument · canvas
Voice, motion, and memory become a field that cannot be screenshotted into the same image twice. An exploration of ephemeral interaction design.
Domains: Aerospace simulation · UAV perception · AI safety & reliability · Uncertainty quantification
Technical: Python · Gazebo · PX4 · Computer vision · WebGPU/WebLLM · Simulation design · Reproducible research protocols
Philosophy: Frozen holdouts · Preregistered gates · Honest uncertainty · Distribution shift awareness
- Aviary Contributions — Performance optimizations merged into OpenMDAO Aviary
- Voyager/Elodin Validation — Early sim-to-sim validation work (simulation-only, not flight-tested)
- SkillsUSA Regional Champion — 3D Visualization & Animation (2026)
Email: suhas.aug20@gmail.com
Links: Portfolio · GitHub · LinkedIn
Affiliation: River Islands High School · Delta College (CS coursework)
Static HTML/CSS/JS portfolio hosted on GitHub Pages. No framework. Clean, fast, accessible.
AegisLand charts and cockpit assets are mirrored from suhaslord/uav-safety-research.
© 2026 Suhas Beemineni · Simulation research, not flight software