Lab

Prototypes exploring how AI can make people genuinely more capable. Each one starts from a research question and ships as something you can actually use.

Nightshift

Full-stack app

A platform for cognitive scientists, psychologists, behavioral researchers — anyone who designs experiments with human participants — to prototype, simulate, and analyze studies in minutes instead of weeks.

Both simulation modes are grounded in established cognitive mechanisms from the literature — latent factor models calibrated to published factor loadings, BFS-computed construal probabilities, Big Five trait covariance. A parametric engine gives you instant, deterministic results for rapid iteration. An LLM-as-participant mode uses identity-driven personas built on those same mechanisms, adding process-level richness — chain-of-thought traces, strategy descriptions, and behavior that emerges from who the simulated person is (a phone-checking 19-year-old, a methodical retiree). Both are validated against published findings (Ho et al. Nature 2022, Lin & Ma NComms).

Design

Conversational AI agent proposes experiment variants with three distinct research personas

Simulate

Both modes grounded in cognitive mechanisms from the literature — fast parametric or rich LLM-as-participant

Analyze

Opus-powered analysis chat with heatmaps, factor analysis, and peer-review simulation

React 19 TypeScript Framer Motion Claude API Vercel Cognitive Science

Why this matters

Both simulation modes encode real cognitive mechanisms — latent factor models, construal probabilities from path salience, trait covariance — not just statistical noise or prompt engineering

LLM personas are built on those same mechanisms, adding chain-of-thought traces and strategy descriptions researchers can inspect

Construal effect replicates at 85% of published human data (Ho et al. Nature 2022) — and the tool is honest about where it diverges and why

Mode

Prototype

An adaptive AI interface that changes how it guides you based on your working style. Built around a cognitive science insight: people approach tasks with systematically different strategies, and interfaces should adapt to that — not force everyone into the same workflow.

React TypeScript Tailwind CSS Framer Motion Cognitive Science

Planner

Structure first

Explorer

Follow curiosity

Driver

Move fast

Coach

Step by step

Four-in-a-Row vs AI

Interactive

Play the game from our NeurIPS 2024 paper on AI planning. This 4x9 board variant (free placement, connect 4 to win) was used to study how AlphaZero and humans develop different planning strategies. The AI uses minimax search — can you beat it?

Game AI Minimax NeurIPS 2024

Four-in-a-Row

4 x 9 board · Place anywhere · From Lin et al., NeurIPS 2024

Your turn — click an empty cell
move 0
You (X)
AI (O)

My PhD studied how people develop skill at complex tasks — running large-scale experiments with thousands of participants across 9 planning games, identifying what makes some people get dramatically better while others plateau.

That research shapes everything I build. Nightshift simulates how diverse humans would experience an experiment. Mode adapts its interface to your cognitive style. The RL work at Amazon trains agents that collaborate with humans, not replace them. The core question: how do we design AI that makes people progressively more capable?