VANTREL

Inspiration

Most people lose money investing not because they pick bad assets, but because of timing and emotion: they buy the top and panic-sell the bottom. The tools they have don't help. Robinhood gives you a buy button. Robo-advisors park your money and never react. Trading bots run one strategy until the market changes and it breaks. We wanted something that trades for a regular person, adapts to the market, controls risk, and explains every move in plain English.

What it does

Vantrel lets you set your risk with a slider, fund an account ($10,000 of paper money), and let AI trade live crypto (BTC, ETH, SOL) for you.

3 trading brains + a Shield. Calm market: Mean-Reversion buys sharp dips. Trending: Trend rides it. Volatile: Breakout buys confirmed moves. Falling: Shield holds cash.
The AI picks the brain. Rust checks every order. Claude reads the market and chooses the strategy. A Rust risk gate enforces your limits on every order, and the AI cannot override it.
Earned risk. The AI can always tighten your limits. It can only ask for more risk, up to the ceiling you set, after it has earned it: 3+ trades in 30 days, 55%+ wins, and not losing today. Otherwise Rust blocks it with a reason code.
Explains itself. Every trade, brain switch and blocked order comes with a one-line reason, on screen and spoken with ElevenLabs.
Verify, don't trust. Every action is SHA-256 hash-chained and anchored on Solana every 5 minutes. Anyone can re-check the track record in their browser.
Watch it survive 2025. A narrated replay of Jan 2025 → today on real daily prices: Vantrel +1.2% vs just holding −27.8%, fees included.
Talk to your portfolio. An MCP server lets you ask Claude about your account. It has no order tools, so it can't trade.

How we built it

Rust (axum, tokio): indicators, regime detection, the brains, the risk gate, the ledger, the API and a live stream. Indicators are parity-tested against our Python backtest on 1,700 days of BTC.
Claude (Anthropic): the regime agent and trade explanations. Falls back to a rule-based mode if the API is down.
ElevenLabs: spoken alerts and the narrated replay, one voice at a time, cached.
Solana (devnet): anchors the audit hash chain.
Auth0 + guest sign-in: Auth0 when configured. Guests get a private random ID per browser, so judges can try it with no account.
React + Vite + TradingView lightweight-charts: the dashboard.
Python: the backtest, AI agents and MCP servers.
Playwright + GitHub Actions: every PR runs Rust tests, end-to-end tests, a production-configured engine and real-browser UI tests before merge.
Docker + Caddy: deployed with HTTPS at getwiththeprogram.study.

Challenges we ran into

Our first strategy lost money. On 4-hour bars, fees ate the small moves (BTC −53%). Switching to daily bars and adding the Shield turned it around.
Keeping the AI honest. An LLM will happily suggest more risk. We made "the AI proposes, Rust decides" a hard rule in code, every rejection a tested reason code, and loosening a limit something the AI has to earn.
Production-only bugs. Sign-in that worked in dev failed on the live server, and an Auth0 setting caused login loops. We added guest sign-in and a CI job that runs the app exactly like production in a real browser.
Narration timing. Getting the voice to follow the chart without overlapping or lagging took a single-voice queue and per-line timing.

Accomplishments that we're proud of

+1.2% vs −27.8% for buy-and-hold through 2025 into 2026 on real prices, and a worst drop of −17% vs −76% over 2022–26 in the backtest.
A safety layer the AI truly can't bypass, with every rule unit-tested.
A track record anyone can verify on a public blockchain.
A full test suite in CI covering the engine, end to end and a real browser, so the live demo doesn't break.

What we learned

Protecting the downside matters more to regular people than chasing upside.
AI is most trustworthy when it advises and hard code enforces.
Test against production configuration, not just dev. That's where our real bugs were.
Be honest about the numbers: the backtest has 98 trades, so we say it's a small sample, and Vantrel lags in strong bull runs because it's built to protect.

What's next for Vantrel

Stocks and ETFs (stock replay is already built on Alpaca data; paper trading is mirrored to Alpaca).
Real-money trading through a regulated broker, with the same Rust gate.
Mobile app with voice alerts.
More brains and a longer, larger live track record before any real-money launch.