SUHAS BEEMINENI · RIHS '29 · SJ DELTA

Perception reliability for autonomy. Simulation-first research.

Independent work on perception overconfidence and calibrated abstention for UAV safety. The holdouts stay frozen, and the failures stay inspectable.

43% → 1% Phase 6B unsafe
safety_acceptance: false
AegisLand fiducial marker geometry, archival still

AegisLand

Claim

Perception reliability for autonomy. Simulation-first research.

Independent work on perception overconfidence and calibrated abstention for UAV safety.

AegisLand fiducial marker geometry from the frozen holdout, archival still
FLAGSHIP · SIMULATION ONLY safety_acceptance: false

Evidence

The exhibit

AegisLand is a simulation-only research cockpit for UAV perception reliability. Not flight software. Not a safety case. An inspectable archive where unsafe rates and abstention behavior stay visible.

This is the flagship work. The holdouts stay frozen, and the failures stay visible.

Method

How it is built

Simulation-first research means the test set is frozen before any tuning. No post-exposure adjustments. The cockpit displays what happened, including what failed.

Each phase locks a holdout set. Results are recorded. Retuning after seeing the holdout is not a result. It is overfitting with extra steps.

Numbers

What the numbers are

Phase 6B unsafe reduction 43% → 1%
V3 arch unsafe reduction 84.2% → 2.4%
Current frontier Phase 11 · P14R
Holdout exposure Frozen

Honesty

What this is not

Status
FLAGSHIP · SIMULATION ONLY
safety_acceptance
false
Flight readiness
Not flight software
Archive mode
Sim-only

Open

What is still open

The research cockpit is live. The protocol is documented. The repository is public.

Selected work

The same question in two domains: when should a system stay quiet?

Then one instrument that only exists while you use it.

Elodin Voyager PR #769 Merged · 2026

Elodin Voyager PR #769

Open source contribution

Merged pull request to Elodin aerospace simulation framework.

Rust Simulation
AbstainBench 30 questions · Local eval

AbstainBench

Language model abstention

30-question local benchmark for answer vs. abstain. No API keys. Hallucination is a scoring category.

Python LLMs
~200 peak users · 2025

StudySync

AI study platform

Peak ~200 active users. AI-powered study tools and scheduling.

React AI
ECHO / FIELD Instrument

ECHO / FIELD

Canvas experiment

Voice, motion, and memory become a field that cannot be screenshotted into the same image twice.

Canvas WebAudio
1st Regional · 2026

SkillsUSA 3D Viz

Competition

1st place regional, 16th California state in 3D Visualization & Animation.

Blender 3D
2nd place · 2025

Python Game Jam

Competition entry

Runner-up in competitive game development jam.

Python Pygame

Wins

Dated, verifiable results

Research, competition, and open source contribution.

AegisLand Phase 6B unsafe reduction 43% → 1% 2026
V3 architecture unsafe share reduction 84.2% → 2.4% 2026
Elodin Voyager open source contribution PR #769 merged 2026
SkillsUSA 3D Visualization & Animation 1st Regional, 16th State 2026
StudySync peak active users ~200 users 2025
Python Game Jam 2nd place 2025

Experience

Building in the world

Internships, open source, and team work.

2026

AI Engineer Intern

Seagulls

Production AI for OpenStage/Seagulls: assistant-integrated exercise search, chat UX, tracing/QA, and Core AI work around routing, memory, safety, and agent architecture.

2026

Fleet Outreach & Business Development Intern

Cruze

Supported fleet outreach and business development focused on moving qualified prospects toward pilot agreements; researched prospective fleet partners and structured outreach around operational value.

2025–26

Volunteer & Program Participation

NSRI · Venture Starters · Learn To Be · STEM·E

Research support, startup mentorship, tutoring, and game development team work.

Now

Student

River Islands HS · Delta College

Class of 2029. College-level CS alongside self-directed work in systems, simulation, and research. San Joaquin Delta College concurrent enrollment.

Contact

Let's build something careful.

Open to research conversations, engineering internships, and collaborations that can survive a holdout.

suhas.aug20@gmail.com