Evidence before advice
Market and portfolio answers use the exact context visible in the workspace.
LIVE PRODUCT · AI × MARKET RESEARCH
Crypto
Copilot
A bilingual AI market-research workspace grounded in live data—not a trading terminal.
01 · THE RESET
From a persuasive concept to a product that can defend every number.
The first version looked like an automated trading platform, but its portfolio, returns, and wallet state were demonstrations. Shipping it meant narrowing the scope and making the remaining loop real.
If an AI finance product cannot tell you where a number came from—or what it does not know—the interface is only performing confidence.
Research before execution
The product helps inspect, compare, and test; it never connects a wallet or places an order.
Unknown stays unknown
No invented news, silent mock fallback, fake returns, or “deployed” strategy states.
7tracked market assets
2real data pipelines
3reproducible backtest rules
2×English + Chinese UI
02 · THE PRODUCT LOOP
Observe the market. Ask with context. Test an idea. Keep the boundary visible.
Four connected workspaces turn a price feed into a research routine without pretending to be an exchange.
LIVE MARKET
A market surface built for comparison, not spectacle.
CoinLore powers the current overview while historical daily candles support 7, 30, and 90-day charts. Normalized comparison makes two different price scales legible.
- Search and filter seven supported assets with visible source freshness.
- Watchlists, saved views, asset groups, notes, and snapshots persist in the browser.
- Portfolio value only appears after the user enters a real quantity.
WORKERS AI · GROUNDED
The AI is useful because its context is constrained.
Each request sends only the question and the necessary market or portfolio summary to Cloudflare Workers AI. The interface shows what the model can see before the user asks.
- Answers cite the current 24H and 7D values available in the product.
- Portfolio analysis remains disabled until real holdings exist.
- There is no scripted answer or silent mock response when the model is unavailable.
REAL PRICE HISTORY
A losing result is better evidence than a fake success state.
The strategy lab tests weekly DCA, daily-dip, and seven-day moving-average breakout rules against real Kraken daily candles.
- 30 or 90-day windows with explicit parameters and signal counts.
- Period return, invested capital, ending value, and maximum drawdown.
- Assumptions state that fees, slippage, taxes, and liquidity impact are excluded.
The screenshot intentionally keeps the negative result. The product reports what happened in the selected period; it does not manufacture proof.
03 · TRUST BOUNDARY
The safest interaction is the one the product clearly does not offer.
What it does
- Reads current market data and historical daily candles
- Calculates user-entered portfolio values locally
- Runs deterministic research backtests
- Sends bounded context to Workers AI for analysis
What it never does
- Connect a wallet or request a signature
- Store private keys or seed phrases
- Place, schedule, or automate trades
- Present research results as financial advice
04 · SYSTEM
A small architecture with explicit seams between data, AI, and local state.
CoinLore + Kraken
Current overview and daily OHLC history.
Pages Functions
Validation, normalization, caching, and safe upstream fallbacks.
React workspace
Market, portfolio, AI, strategy, and bilingual UI.
Workers AI + localStorage
Bounded AI requests; holdings and research remain local.
React 19TypeScriptViteZodCloudflare Pages FunctionsCloudflare Workers AICoinLoreKrakenlocalStorage
05 · WHAT CHANGED
Trust did not come from adding more AI. It came from removing claims the product could not yet support.
Scope is a design material
Removing wallet and execution flows made the research loop clearer and the privacy story defensible.
Grounding needs an interface
Showing source freshness and “what the AI can see” turns a backend constraint into user confidence.
Failure must remain legible
Negative backtests, missing context, and provider errors stay visible instead of becoming polished fiction.