Educational Real-World PoC · Live Paper Trading

A multi-agent AI trading system

Five specialist LLM agents propose trades, a critic validates them, my phone gets a veto window, and a deterministic risk guard has final veto — with a local NVIDIA DGX Spark reading the whole market every morning. Everything below is a live feed from the system running as an educational real-world proof-of-concept on a paper account.

Paper trading only — for research and demonstration. Not investment advice.

LIVE

The strategy — how & why

It's not one bet — it's a diversified ensemble of five classic equity strategies run in parallel on the S&P 500, long and short, on a multi-day (swing) horizon. The key idea: the edge is not an AI predicting prices. It comes from two things a human desk can't match at scale.

1 · Information breadth

A local NVIDIA GPU reads the entire universe every day — news, 8-K filings, social chatter, analyst revisions, search-attention spikes — and scores it. The agents start each morning having "read" far more than any analyst could.

2 · Feedback loops

Every closed trade is attributed to the agent that made it. Capital is re-weighted toward whichever strategies actually earn it, and a RAG memory recalls how similar past setups resolved. The system is built to measure edge, not assume it.

The five specialist agentsi

Momentum Trend & relative strength

Longs breakouts backed by volume; shorts relative weakness / broken uptrends.

Macro Sector rotation & factors

Longs names in regime-favored sectors; shorts laggards in out-of-favor ones.

StatArb Mean-reversion pairs

Shorts the rich leg and longs the cheap leg of a diverged, correlated pair.

Contrarian Oversold reversal

Buys irrational panic; shorts broken theses (high-severity news / downgrades).

Exotic Event-driven catalysts

Longs positive catalysts (beats, raises); shorts misses and guidance cuts.

Crypto 24/7 spot book (separate account)

Long/flat swing trades on ~30 major pairs, stops sized at ~2σ of each pair's own volatility.

Insider (shadow) Form 4 conviction — research lane only

Officer/director open-market buys, cluster-weighted, 10b5-1 plan trades discounted. Trades zero capital until 20+ settled shadow trades prove a positive expectancy CI.

Every trade clears four gatesi

1 · Propose

Five isolated agents each read a different slice of the market and emit structured trade ideas.

2 · Critic

An LLM critic scores every idea against an investment memo and kills the weak ones (~60%).

3 · Announce + veto

Every survivor is auto-approved and announced to my phone with a veto button — I can stop any trade before it executes, without being a bottleneck.

4 · Risk guard

A pure-Python, no-LLM guard has final veto: 10% position / 30% sector / 1.0× gross. Then capital flows to whichever agents actually earn it.

Why this way: diversified strategies + hard risk limits + measured attribution means no single idea can sink the book, and the data tells us which approaches have edge instead of us guessing. Gross leverage is capped at 1.0× on purpose — the goal right now is to prove edge, not amplify an unproven one. That's what the scorecard below is for.

Live system feedi

loading…

Equity curve (paper, $25k start)i

Governance funneli

How many ideas survive each gate from proposal to execution.

Per-agent realized P&Li

Alpha factory throughputi

Component healthi

DGX Spark — live GPU loadi

The local NVIDIA GB10 (spanky1) running vLLM — it classifies and embeds the market's news & filings all day. This is its real workload, scraped from the model server every minute.

Throughput — generation tokens / sec (last 6h)

What the system is tradingi

No black box — these are the actual open paper positions and recently closed trades, straight from the live feed. Equities only (long/short bracket orders); every one cleared the critic, a human veto window, and the deterministic risk guard.

Open positions

TickerSideEntry TargetStopAgent

Recent closed trades

TickerSideEntry ExitP&LWhy

Paper trading — for research and demonstration. The account identifier and order IDs are never exposed.

⚡ Intraday book i

Deterministic opening-range breakouts on the ~150 most liquid S&P names, flat by close every day. Its job is velocity of evidence: same-day settlement means the n≥50 verdict arrives in weeks. Judged on its own numbers — never pooled with the swing books.

How trades ended (fill-based only) i

ExitCountP&L

Recent intraday trades

TickerSideEntry ExitP&LWhy

🛞 Wheel book i

Open option legs i

ContractPremiumClose at

Resolved legs

ContractP&LHow

🪙 Crypto book i

Runs on its own account and its own clock — same five gates (propose → critic → veto window → risk guard), long-only, exits managed in software.

Crypto equity curve i

$15,000 start · marked 24/7 · dashed line = break-even

Open crypto positions

PairSideEntry TargetStop

Recent crypto trades

PairEntryExit P&LWhy

Does it actually have edge?i

A trading system that shows a win rate without a confidence interval is lying to you. This scorecard measures whether closed-trade returns are statistically distinguishable from zero — and refuses to claim edge until the sample is big enough.

Per-agent edgei

AgentClosedHit rate (95% CI) Expectancy/tradeTotal P&LProfit factor

Live system feed EVERY OWNER MESSAGE i

The same messages that reach the owner's phone, unedited: proposals, verdicts, exits, alerts, recoveries.

🧪 Research lane HYPOTHETICAL — NO CAPITAL i

Before any strategy idea touches money, it runs here: real bars, real fees, zero capital. Eight signal families and counting — stop geometry, insider flow (SEC Form 4), social momentum, hour-scale breakouts and panic-dip reversion on minute bars, weekly reversal, turn-of-month — each with matched controls where beta could masquerade as edge, and a multiple-testing bar before anything is promoted. Most cells measure NEGATIVE or flat, which is the point: every cheap no permanently closes a temptation, and the rare near-positive (turn-of-month, 7 years of month-ends, CI a hair under zero) earns patient monitoring instead of blind capital. These numbers are simulations and are never blended with the live books above.

CohortSettledStill open Win rateStopped outVerdict

💰 Institutional money flow i

Where the big orders went today — net $ millions by order-size class, across the whole S&P 500. Context the agents read, not holdings.

Top institutional inflows

SymbolInstitutional netRetail net

Top institutional outflows

SymbolInstitutional netRetail net

Methodology: hit rate uses the Wilson score interval; per-trade expectancy is the mean return on deployed capital with a 95% t-interval. Paper trading — not investment advice, not a track record.

How this live feed reaches the page

The system runs entirely on my own hardware — there's no public database. Here's how I safely surface a live view of it on this public site:

1 · Postgres

The same database that powers my internal Grafana dashboards stores every trade, snapshot and equity point.

2 · Metrics API

A tiny read-only service publishes only curated aggregates — equity curve, returns, per-agent P&L, health. No account number, IPs or secrets.

3 · Cloudflare Tunnel

A zero-trust tunnel exposes that one endpoint at trading-api.vitalemazo.com — no ports opened, nothing else on my network reachable.

4 · This page

Your browser fetches that feed directly and re-renders every 60 seconds — so the numbers above are always live.

Internal observability (Grafana, alerting, the auto-healing health agent) stays private behind Cloudflare Access — only the read-only aggregate feed is public.

How it works

The design principle: agentic models reason, the local GPU grinds volume. The edge comes from processing the entire market cheaply and locally, plus tight feedback loops — not from an LLM guessing stock prices. Every trade passes through a human and a deterministic guard.

System architecture

Infrastructure & delivery

This is an educational real-world proof-of-concept — a full platform-engineering pipeline running on my own hardware: GitOps deploys, self-hosted compute, secret management, and zero-trust exposure.

Infrastructure and delivery flow

Git as the control plane

The unusual part of this platform isn't that it deploys from git — it's who commits. Two autonomous AI agents work on this repo around the clock: a maintenance agent that finds and fixes bugs every six hours with full merge authority, and an independent supervisor that audits the first agent's work twice a day. Both ship through exactly the same gate a human does.

The repo is the constitution

Two charter files govern the agents: one defines what they may change alone (bugs, observability, research) versus what always needs a human (risk limits, money movement); the other gates what the system may become on measured evidence. Agents read these at the start of every run — the rules version with the code they govern.

Merge = production deploy

A push to main runs lint plus 590+ tests, then a self-hosted runner rebuilds the image and recreates all 19 service containers — re-running the full suite first. The deploy log echoes every container back up with its exact command, because a green workflow only proves the script exited; the roll-call proves the fleet is actually running the new code.

Strategy proves itself first

No strategy change goes straight to money. Variants run in the shadow research lane — same setups, real bars, real fees, zero orders — and promotion to live requires 20+ settled trades with a positive expectancy confidence interval, filed as a pull request with the evidence in the body. The audit trail of an AI-operated trading system is just git log.

Git as the control plane: humans and AI agents shipping through one gated pipeline

Built

Data layer

Nightly S&P 500 snapshot — Alpaca OHLCV + Benzinga news, SEC EDGAR filings, FRED macro — into Postgres.

Alpha factory (DGX)

A local NVIDIA GB10 classifies every news article & filing and embeds them for RAG memory, plus an O(N²) pairs scan.

Five specialist agents

Momentum, Macro, StatArb, Contrarian, Exotic — isolated and parallel, each an agentic model reasoning independently, emitting structured JSON trades.

Critic + human gate

A critic validates against an investment memo; every trade is announced on Telegram with a human veto window before it executes.

Deterministic RiskGuard

A pure-Python, unit-tested guard has final veto: 10% position / 30% sector / 1.0× leverage. No LLM in the loop.

Feedback loops

Per-agent P&L attribution → dynamic capital weights, plus a self-healing health agent and full Grafana observability.

Research lane (Stage 0R)

New strategies prove themselves on real bars with zero capital before touching money — including a deterministic insider-flow agent built from SEC Form 4 filings, entered only at prices a filing reader could actually have obtained.

Autonomous maintenance

Two cloud AI agents work the repo around the clock — one fixes bugs with full merge authority inside a written charter, one audits it independently. Every change ships through the same gated pipeline a human uses.

Stack

Python · uvPostgres + pgvectorvLLM · Qwen NVFP4Agentic reasoning modelsAlpaca (paper)Docker · UnraidHashiCorp VaultGitHub ActionsTerraform · CloudflareGrafana