HoloTrade: AI-Powered Financial Learning Platform

Inspiration

Most of our friends have never bought a stock. The few who tried gave up in ten minutes — not because investing is hard, but because every platform shows a candlestick chart and assumes you already know how to read one. We wanted to build the missing first step: a teacher sitting next to you while you look at the chart.

What it does

HoloTrade Mentor is Yahoo Finance with a teacher inside it. It shows live candlestick charts for Apple, Microsoft, NVIDIA, and Tesla — paired with Holo, a Claude-powered AI mentor that explains what's happening in plain English. Open a chart, Holo introduces it. Switch the interval, Holo re-explains. Ask "what does this dip mean?" and Holo answers using the actual numbers on screen. Ask "should I buy this?" and it refuses, then pivots to teaching. No trading, no portfolio, no real money — just a stock chart turned into a classroom.

How we built it

Frontend: React + TypeScript + Tailwind, with TradingView's lightweight-charts. Backend: FastAPI in Python, proxying Twelve Data (market data) and the Claude API. Dual-model strategy: Claude Opus 4.7 for deep auto-explanations on chart load, Claude Sonnet 4.5 for fast streaming Q&A.

Chart context injection: before every Claude call, we wrap the recent OHLC candles, trend, and volatility in an XML block prepended to the prompt — so Holo grounds every answer in real prices, not generic phrases. Streaming end-to-end via Server-Sent Events, with TTL caching in front of Twelve Data so the demo never hits rate limits. Deployed on Vercel (frontend) + Render (backend).

Challenges we ran into

Keeping Claude grounded in real numbers. Our first prompt produced beautifully written but generic answers. Pre-computing trend and volatility on the backend, then forcing the model to reference specific prices, fixed it. Closing the advice loophole. "Never give advice" wasn't enough — Claude still hedged with phrases like "Tesla looks strong." We rewrote the prompt with explicit forbidden phrases and a worked refusal example. Now Holo redirects every time. Streaming through the full stack. Getting tokens to flow smoothly from Claude → FastAPI → React without buffering took disabling proxy buffers, escaping newlines in SSE frames, and using fetch + ReadableStream instead of EventSource.

Accomplishments that we're proud of

Holo's explanations actually reference real prices — "NVIDIA opened at $X, currently $Y, a Z% move" — not generic descriptions. The ethics layer is unshakeable. We tested every trick to get advice out of Holo. It refuses every time. The product feels alive. Streaming tokens, pulsing avatar, auto-explanations on chart change — it feels less like a hackathon demo and more like a real product.

What we learned

Prompt engineering is real engineering. We rewrote the system prompt eight times, each iteration closing a specific failure mode we'd watched happen. Right model for the right job. Opus for depth, Sonnet for speed — using both made the demo feel dramatically better than picking one. Cut features to win demos. We shipped four features well instead of six poorly. Judges score what they see in two minutes.

What's next for HoloTrade: AI-Powered Financial Learning Platform

Paper trading simulator — practice with virtual money, no real risk. Personalised learning paths — Holo remembers what concepts you've mastered and what to teach next. Risk analysis engine — beginner-friendly risk scores framed as education, not as buy/sell signals. Voice-enabled Holo — spoken explanations for accessibility. More markets — ETFs, mutual funds, eventually crypto.

Built With

• antropic
• api
• claude
• python
• typescript