Aditya Rao

About

I build agent tooling and LLM evaluation, and I've worked on production ML systems.

My first research project was a gesture-recognition system to replace the iClicker in lecture halls. It looked at what a camera needs to poll a real class: telling apart hands that block each other, keeping up in real time without lag, and the privacy of filming a room full of students.

I red-teamed LLMs for 3 international businesses that wanted to know how their own tools could be broken. I've pitched proof-of-concepts to C-level executives. And the NBA model beat ESPN's pregame odds across 400+ games.

I co-founded PlaySmart Academy, which teaches K-5 math through live sports. By day I'm a software engineer at a Series B fintech, building the agent harness behind its AI assistant.

I've also built fall detection from wearable sensor data, a Roblox game in Lua, and a Flask tool that cut a YouTube channel team's transcript formatting by about 80%.

Outside of work I follow sports closely, and I like to read about healthcare and space.

Research

Most of my research tests one thing: whether a model still works once the data stops looking like what it was trained on.

Outside of that, I work on LLM agents and how to evaluate them, forecasting, and vision on real-world devices.

Robust Graph-Based Android Malware Classification Under Distribution Shift

San José State · Spring 2026

Cut the robustness gap from 36.2 to 31.0 points.

How far does a graph neural network's accuracy fall when Android malware stops looking like its training data, and does telling the model which features are missing close that gap? Training with simulated missing features cut the robustness gap from 36.2 to 31.0 points. The gating model I proposed did not beat that, most likely because the extra features were synthetic.

Bar chart of standard against shifted test accuracy for four graph models. Each scores about 72 to 75 percent on the standard split and 38 to 43 percent on the shifted one.

Identifying Subgiant Stars and Binary Systems in Gaia DR3

UC San Diego Stellar Data Lab · advised by Prof. Chris Theissen

Narrowed 288,746 stars to about 53,800 subgiant candidates.

I queried Gaia DR3 and cleaned the sample to 288,746 stars. Using temperature, surface gravity and absolute magnitude, I isolated about 53,800 subgiant candidates. I cross-matched them against the El-Badry wide-binary catalog to find subgiants with white dwarf and low-mass companions. I presented the work at a research symposium at UC Irvine in October 2025.

Kiel diagram of the Gaia DR3 sample, surface gravity against effective temperature, with the giant, subgiant and main-sequence regions labelled.

S&P 500 Forecasting Using Generative AI

Peer-reviewed · UC Merced Undergraduate Research Journal

A GPT-4-based model for short- and long-term S&P 500 forecasting, improved iteratively. GPT-4 was new at the time, and the question was what it could actually do.

V-QUIP: Vision-Based Quick Impromptu Polling

Google exploreCSR

30% lower latency.

Fine-tuned YOLO and optimized TensorFlow Lite models for real-time gesture recognition in classrooms, cutting latency by 30%.

Fall Detection for Rare Events and Changing Devices

Guardian AI

Judged only on people held out of training.

A missed fall costs more than a false alarm. Falls are rare in the data, and sensors differ between devices. I trained CatBoost on 3-second windows of SisFall accelerometer and gyroscope data. I balanced the fall class with ADASYN and mixup, and judged the model only on people held out of training. Temporal smoothing cuts false alarms, and a CORAL alignment step carries the model from waist-worn sensors toward wrist devices.

Identifying Early-Stage Alzheimer's Disease from MRI

Early research

Early research, begun before ChatGPT existed: a TensorFlow/Keras convolutional network that sorts brain MRI scans into stages of Alzheimer's disease, trained with data augmentation, SGD and a learning-rate schedule.

More projects

  • MapnosisMar 2024

    Code ↗

    Winner, Best Use of MongoDB Atlas · HackMerced IX

    A site that maps pollen and air-quality risk for any location. It gives forecasts and risk advice for the health conditions a user selects, and it has a chatbot on environmental pollutants. Auth0 handles login, and MongoDB Atlas stores each user's preferences.

  • BBALL HQ Discord botApr 2024 - May 2024

    Code ↗

    A Discord bot for NBA fans. It shows live scores behind spoiler tags, live play-by-play, shot charts and heatmaps for players and teams, stats, and news.

  • Motion detection systemMay 2024

    Code ↗

    An M5GO device with a PIR sensor that sends a text alert when it detects motion, with a live on-screen clock whose colors shift between day and night.

  • Mask approval checkerJun 2022 - Jun 2023

    code not public

    Sorted FDA mask approval IDs by status and used Amazon's shopping API to flag listings selling unapproved protective masks.

Career

Eighteen roles since 2022. The first were student clubs and research. The most recent is a full-time engineering job.

2026

  • Series B Fintechnow

    Software Engineer

    Building the agent harness behind a consumer product's AI assistant, from orchestration through evaluation

    Jun 2026 – Present

  • Stealth AI Startup (a16z-backed)

    Founding AI Engineer

    Wrote LLM-powered CRM and outreach features as founding engineer

    Feb 2026 – Jun 20265 mo

2025

  • Guardian AI

    Co-Founder & CTO

    Trained fall detection on wearable sensor data, designed to move from waist to wrist sensors

    Aug 2025 – Jan 20266 mo

  • Cybersecurity Startup (Series B)

    Machine Learning Engineer Intern

    Built top-level domain ingestion pipelines with ML

    Aug 2025 – Dec 20255 mo

  • Sixfold AI (Series B)

    Full Stack AI Engineer Intern

    Developed a prompt-refinement and LLM evaluation tool on Langfuse traces

    Jun 2025 – Aug 20253 mo

  • PlaySmart Academynow

    Co-Founder & CTO

    3 California pilot schools committed, with 14-minute average sessions at 2.5× the edtech baseline

    May 2025 – Present

2024

  • Independent Client

    Software Engineering Consultant

    Cut a YouTube channel team's transcript formatting by ~80%

    Nov 2024 – Mar 20255 mo

  • HoopSight AI

    Founding Machine Learning Engineer

    Beat ESPN's pregame odds with an NBA model at 74% across 400+ games

    Aug 2024 – Feb 20257 mo

  • Model N

    Machine Learning Product Intern

    Cut quote-creation turnaround 30% with supervised ML, finishing in 8 weeks a project planned for a full fiscal year

    Jun 2024 – Aug 20243 mo

  • Pick N Show

    Full Stack Developer

    Created a photographer-to-client platform with recommendations

    Feb 2024 – Jun 20251 yr 5 mo

  • Upwork (freelance)

    LLM Evaluation Specialist (Red-Teaming)

    Red-teamed LLMs for 3 international businesses

    Jan 2024 – May 20245 mo

2023

  • ChatBuilt

    AI Chatbot Developer

    Lifted retention 40% in beta with a GPT-4 Turbo chatbot

    Dec 2023 – Jan 20242 mo

  • Finance-Flash (Alphashots.AI)

    ML Researcher

    Researched stock-prediction algorithms and pattern matching

    Dec 2023 – Feb 20243 mo

  • Google exploreCSR

    Undergraduate Researcher

    Built an iClicker alternative driven by gesture recognition

    Oct 2023 – Dec 20241 yr 3 mo

  • Open Avenues

    Signal Processing & CV Engineering Intern

    Recognized text in natural images with the EAST detector Code ↗

    Sep 2023 – Nov 20233 mo

  • DIA Higher Ed

    AI Implementation Intern

    Demonstrated sentiment analysis to help colleges read student surveys

    Sep 20231 mo

  • ACM UCM

    Data Science Lead

    Ran nine data science workshops, from NumPy and SQL to deep learning in PyTorch

    Aug 2023 – May 202410 mo

2022

  • The Statistician

    Communications Manager & Data Analyst

    Analyzed the NBA playoffs in TensorFlow and grew the club's Discord past 50 members

    Jul 2022 – May 202311 mo

Tech stack

Programming Languages

  • Python
  • SQL
  • TypeScript

LLM Eval & ML Frameworks

  • PyTorch
  • scikit-learn
  • Braintrust
  • Langfuse

Full-Stack & Product

  • Next.js
  • React
  • GraphQL
  • Redis
  • Supabase

Data & Analysis

  • PostgreSQL

Deployment & MLOps

  • Docker

Education

San José State University

The Trustees of the California State University
on recommendation of the faculty of the

College of Science

have conferred upon

Aditya Rao

the degree of

Bachelor of Science

Computer Science

Cum Laude

Honors in Computer Science

with all the rights and privileges pertaining thereto,
given at San José this twentieth day of May, two thousand twenty-six.

Languages: English (native), Spanish (limited working proficiency).