Aidan Whitedeer
CS Undergraduate at Stanford and Researcher at Stanford Virtual Human Interaction Lab. Previously at Microsoft, building at MTS Live, incoming at Uber AI. Research interests include virtual spaces, human interaction within VR/XR spaces, and reinforcement learning.
Featured Projects
Parallel Deep-RL Agents for Roblox Obstacle-Course Navigation: From Single-Course Memorization to Generalizing Across Procedurally-Composed Courses
Department of Computer Science · Stanford University · 2026
A shared deep-RL agent (Behavioral Cloning → PPO) trained inside Roblox to navigate platforming obbies, with many agents running inference in-engine in parallel. We build a full pipeline that bridges Luau game physics to a training loop, procedurally generates obstacle courses, and trains a single course-agnostic policy across them via domain randomization.
Sentiment-Augmented Cross-Sectional Stock Selection
Department of Computer Science · Stanford University · 2026
We investigate whether sequentially filtering cross-sectional equity signals (first by traditional factors, then by FinBERT sentiment) outperforms joint modeling, and identify the conditions under which sentiment adds conditional predictive value within a factor-screened universe.
Experience
We built a way for you to preview ElevenLabs' new musicgen model, Music V2, along with two audio models: Voice and Scribe.https://t.co/oIQ8cpYZp8 https://t.co/7jkh2T0tLz pic.twitter.com/MVMJSivNAB
— MTS (@MTSlive) June 1, 2026
OpenAI and Anthropic's Mythos just independently disproved Erdös' unit-distance conjecture six days apart.
— MTS (@MTSlive) May 28, 2026
We broke down the math, the models, and what it means that machines are now formally proving theorems.
Read the full drop: https://t.co/neBOdAKu3J pic.twitter.com/UQXR0MP1cl
Personal agents have gone mainstream.
— MTS (@MTSlive) June 16, 2026
Always-on software that runs on your machine: it reads your messages, writes its own code and handles tasks on its own.
We broke down the loop, the projects, and what exactly makes them so useful for everyday work. https://t.co/Ilenw5celC pic.twitter.com/0xAgVF43km
Interesting. https://t.co/Vt7fvjrShj
— Marc Andreessen 🇺🇸 (@pmarca) May 25, 2026
Featured Projects
A test-time-training loop that finetunes a config-generating LLM via LoRA (SFT/DPO) on high-reward (prompt → hardware config) trajectories, with reward computed from bit-accurate C-simulation and Vivado synthesis. Engineered lexicographic, budget-normalized rewards over a precision × reuse × strategy search space, plus an agentic repair loop that recovers from compile failures. Beats random-search, surrogate, and frozen-LLM baselines.
An AI-powered DJ system with dual-deck audio engine, BPM detection, automated transitions, 3-band EQ, and bass/voice/melody isolation. Built with Next.js 16, React 19, TypeScript, and Grok/X API integration. Features a real-time 3D visualizer with Three.js/React Three Fiber, voice control with speech recognition and TTS, and AI copilot music automation.
An AI-powered SDR system with automated lead scoring and pipeline management. Built automated lead scoring, email generation, and pipeline management with REST APIs, multi-model evaluation, error handling with retries, and production-ready architecture.
An AI-powered Web3 payment system built with TypeScript/Next.js, Coinbase CDP wallets, Claude agents, and Vapi voice calls. Supports autonomous transaction flows, HTTP 402 payments, cost-based tool chaining, phone approvals, and gas-abstracted on-chain payments using the Coinbase + Anthropic SDK.
Creating framework for tribes to finetune mBERT models for endangered Native American Languages. Working with Raices Collaboration Project, mentored by Ken Attocknie. Received Dreamstarter Grant for $20,000 towards the project.