Live · Paper Account

Quantis AI Trading

An autonomous AI agent runs disciplined swing-trading research on a live US-stock paper account. Five scheduled jobs per weekday. Hard rules the LLM cannot override. Every decision committed to git.

Personal engineering project · Paper trading only · Not investment advice

What this is

A discretionary, catalyst-driven swing-trading bot for US equities. An LLM reads news, identifies setups, applies a fixed rule layer, and places paper orders. The strongest part of the design isn't the alpha — it's the discipline. Trailing stops fire automatically. Losers get cut at –7%. Sectors with two consecutive failed trades get exited. Patience over activity, every time.

One trading day

How it works

◐

06:00 CT

Research

Reads market context — oil, S&P futures, VIX, today's catalysts, earnings calendar, sector momentum, news on every held position — via Perplexity. Drafts trade ideas with explicit catalyst, entry, stop, and target.

▲

08:30 CT

Execute

Re-validates planned trades with live quotes, runs an 8-check buy-side gate (position count, weekly trade cap, sizing, PDT, drawdown circuit-breaker), executes the buy, and immediately places a 10% trailing-stop GTC order on Alpaca.

◑

Throughout day

Triage & review

Midday cuts losers at –7%, tightens trailing stops on winners (+15% → 7%, +20% → 5%). EOD writes a portfolio snapshot. Friday's weekly review uses Claude Opus to synthesise stats vs SPY and grade A–F.

Stack

How it's built

No databases, no ORM, no in-memory state. Every memory file — strategy rulebook, trade ledger, daily research, weekly reviews — is a markdown file committed to a private GitHub repo. Every routine run is a fresh container that clones, decides, commits, exits.

Claude Sonnet 4.6

Daily research & execution

Claude Opus 4.7

Friday weekly review

Alpaca

Paper brokerage

Perplexity Sonar

Real-time research API

Next.js + Vercel

This dashboard

Git as memory

Every decision committed

The rulebook

Hard rules the LLM cannot override

Strategy discipline is enforced before the order is placed, not during reasoning. The LLM proposes; the rules dispose. Every buy must pass an 8-check gate — position count, weekly trade cap, sizing, available cash, PDT room, documented catalyst, ticker validity, drawdown circuit-breaker.

Stocks only — no options, ever
Max 5–6 open positions, ≤20% per position
Max 3 new trades per week
10% trailing stop on every fill (real GTC order — never mental)
Hard cut at –7% from entry
Tighten stops on winners (+15% → 7%, +20% → 5%)
Halt new entries if account drops 15% from peak
If 30-day return underperforms SPY by 5%+, weekly review must consider strategy halt

Honest expectations

What this actually is

~15–25%

Probability the bot beats SPY (S&P 500 ETF) over 6–12 months. Base rates for retail active strategies are not kind — this strategy isn't magically better than average.

~80%+

Probability the bot doesn't blow up. The strongest part of the design is the trailing-stop + hard-cut discipline. It won't make us rich, but it won't destroy capital either.

This is a learning project. The infrastructure (autonomous LLM agent, hard-rule discipline, git-as-memory) has engineering value regardless of P&L. There is no signal service, no subscription, nothing for sale, no community — just a private dashboard for the operator and a public page explaining what was built.

See it run

The dashboard is private. The only person with access is the operator. Click below if that's you.