Playground

All Projects & Experiments

  • ◇ EXPERIMENT

    AI AGENT · COST & EVALUATION

    RepoCoach

    Two paper estimates, both wrong — only measurement got it right. A vertical slice of a source-reading agent whose real output is data on the counter-intuitive parts of agent engineering: cost grows super-linearly with call count, 28% fewer tool calls raised tokens by 17%, and a 65.6% cache-hit rate inflated the budget metric roughly threefold. It also pinned down how safety gates have to be checked in both directions, how state survives across processes, and why a constraint written only into a prompt still needs a gate at the exit. 638 tests, 22 merged PRs — planned by Claude, implemented by DeepSeek, reviewed by GPT. Development has since been wound down, with the measured parts written up as reusable findings.

    GitHub →

  • ◇ EXPERIMENT

    AI COST · MEASUREMENT

    Agent Cost Lab

    Tokens are not money. Nearly every context-compression tool advertises a token-reduction number, but a cached input token costs ~10% of list price and compression works by rewriting history — which breaks the cached prefix and reprices everything after it. "40% fewer tokens" and "a bigger bill" are entirely compatible. This repo measures the bill: a proxy probe on the real call path, predictions locked into git before measuring, export gates written as tests rather than checklists. Instrument built; results pending.

    GitHub →

  • ◇ EXPERIMENT

    AI AGENT · DESIGN TO CODE

    Image to Code

    Not "a pixel-perfect page from one screenshot". A Claude Code workflow that turns AI-generated design sources into reusable, validated frontend packages — preserving real assets and design language, materialising component files with props, variants, tokens, and manifests instead of flattening everything into one page component. Observable intermediate artifacts, deterministic contracts, cheap validation, and human review only where visual judgment is still necessary.

    GitHub →

  • ◇ EXPERIMENT

    AI DESIGN · RESEARCH

    Design Reasoning Lab

    Not "what can this tool generate" but "how does it think". Reproducible behavioural experiments reverse-engineering how AI-native design tools reason — what they ask the human for, which decisions they delegate, and how output modes relate. The current case study is Claude Design with v0 as contrast; claims are kept at four distinct epistemic levels — fact, interpretation, hypothesis, conclusion — so observation is never merged into explanation.

    GitHub →
  • Brooch Shop

    ● LIVE

    AI AGENT · E-COMMERCE

    Brooch Shop

    Where filters fall short, agents begin. An AI shopping agent built with Claude API and tool calling — demonstrating how natural-language intent understanding replaces multi-level filter UX.

    GitHub →Live Demo →
  • Prompt Shop

    ● LIVE

    CHECKOUT · STRIPE

    Prompt Shop

    End-to-end e-commerce checkout playground with Stripe integration — covering the full flow from product listing to payment confirmation.

    GitHub →Live Demo →
  • Human vs AI Judge

    ● LIVE

    AI · GAME · E-COMMERCE

    Human vs AI Judge

    When AI has already decided, who presses confirm? Compete with AI on intent classification of real e-commerce support tickets — compare human, Claude, and GPT judgments in real time.

    GitHub →Live Demo →
  • Readable & Shareable

    ● LIVE

    CLAUDE CODE · SKILL · VISUALIZATION

    Readable & Shareable

    You wrote a great article — most people won't click the link. A Claude Code `/viz` skill that detects article structure and generates mind maps, interactive HTML, and shareable image cards — each output links back to the source so every audience finds their entry point.

    GitHub →Live Demo →
  • Particle Flow Generator

    ● LIVE

    CANVAS · SIMULATION

    Particle Flow Generator

    Stacking time into space. Simulating the spatial residue of user trajectories with particle flows — Canvas 2D real-time rendering, BFS flow-field pathfinding, live parameter tuning and GIF export.

    GitHub →Live Demo →
  • SSB Website

    ● LIVE

    DESIGN ENGINEERING · FRAMER MOTION

    SSB Website

    A design engineering exercise — four-page B2B marketing site built from scratch. Features a custom physics bounce hook with velocity vectors and wall-reflection, responsive layout, and Framer Motion animations. Lighthouse: Performance 91 / Accessibility 96 / SEO 100.

    GitHub →Live Demo →
  • Design Token Signals

    ◇ EXPERIMENT

    DESIGN SYSTEM · EXPERIMENT

    Design Token Signals

    Design tokens don't just store values — they carry meaning. An experimental tool for analyzing the visual signals that token combinations transmit: mapping the semantics of color, radius, and weight to surface conflicts before a component is built.

    GitHub →Live Demo →
  • EaseBuy

    ◆ HACKATHON ★ TOP 10

    UX DESIGN · E-COMMERCE · MOBILE

    EaseBuy

    Top 10 in UXcel × UX Pilot Design Competition. A minimalist mobile e-commerce app covering the full shopping journey across 11 screens — generous whitespace, card-based IA, transparent checkout. Full flow from wireframe to hi-fi with AI assistance.

    View Submission →
  • Spec-Driven Scenario Updater

    ◆ HACKATHON

    AI AGENT · GITLAB HACKATHON

    Spec-Driven Scenario Updater

    GitLab AI Hackathon submission. Targeting the "spec gap" — the distortion that accumulates as business logic is translated across product, design, and engineering. A single spec source drives AI to auto-generate code, tests, and docs.

    GitLab →Watch →View Submission →