PAPER TRADING ONLY · NO LIVE ORDERS ARE EVER SENT

About

What VOLMIND is, how it decides, and where it’s headed.

Alpaca AI Trading Agents Hackathon

VOLMIND

Trade the divergence between AI belief and market belief.

An autonomous, multi-agent options-research and trading terminal built on Alpaca — and a decision-support instrument a real desk could run alongside its existing tools.

volmind — live pipeline
$ scan DIS
news_agent done confidence 0.55
fundamental_agent done confidence 0.55
probability_agent done P(up) 0.58
divergence_agent computed AI 0.58 vs market 0.50
diligence_agent cleared concern score 0.20
risk_agent approved within limits
execution submitted paper order
for Alpaca 🦙
01Why this exists

Most “AI trading agents” stop at an opinion.

The hackathon brief asks for an agent that identifies opportunities, makes decisions, manages positions, and performs — not one that just narrates a hunch. VOLMIND’s bet is narrower and more falsifiable than “the model thinks it’ll go up”: it forms an independent probability estimate, compares it to what the options market is already pricing in, and only treats a large, well-evidenced gap between the two as a signal worth acting on.

A single LLM call agreeing with itself isn’t a strategy. So the gap has to survive scrutiny before it becomes a trade: an independent second review that can kill the idea, a risk gate with hard dollar and position-count limits, and — the part most weekend projects skip — a monitor that actually closes the position later and turns it into a realized number.

02How it works

Two gates stand between a hunch and a paper order.

Three agents each form an independent view from live Alpaca data. Their combined belief is compared against the market’s own option-implied probability. What survives has to clear diligence — an adversarial second read — and then risk — hard dollar and position limits — before a single order reaches Alpaca.

SIGNAL FORMATIONREVIEWEXECUTION & CLOSE-OUTNEWSFUNDAMENTALPROBABILITYforms AI beliefLIVE OPTION CHAIN · no LLMMARKET PROB.delta-approximatedDIVERGENCEAI − marketDILIGENCEmateriality reviewflagged · no tradeclearedRISKposition & $ limitsvetoedEVALUATORalways runs last · logs whyapprovedEXECUTIONpaper order · Alpacaon a schedulePOSITION MONITORmarks to markettake-profit / stop-loss / expiryAGENT REPUTATIONwin rate per agent
Diligence is the one true branch — a flagged thesis skips straight to Evaluator, never reaching Options Architect. Risk is checked one step later, inside Execution itself: it can still keep an order from ever reaching Alpaca. Position Monitor runs separately, on its own schedule, and its close-out feeds each agent’s win rate back into Agent Reputation.
03What makes it different

Built to be distrusted by its own operator.

Every mechanism below exists because the obvious version of this project would have quietly skipped it.

FAIL CLOSED

Paper, always

Execution refuses to run unless ALPACA_PAPER_TRADE is the literal string "true", checked on every order — opening and closing. The Alpaca client is separately hardcoded to paper mode as defense in depth.

MATERIALITY, NOT THEATER

A gate that can say yes

The diligence reviewer weighs whether a concern is severe enough to erase the edge — not whether it can find one. Every real option carries some risk; finding one isn't sufficient grounds to kill a trade.

PORTFOLIO AWARE

Risk beyond one trade

Per-trade max loss, a cap on concurrent open positions, and a daily realized-loss circuit breaker — a bad run stops the agent from digging in further the same day.

CLOSES THE LOOP

P&L, not just picks

A Position Monitor marks every open trade to market and force-closes it on schedule. That's what turns a signal into a number, and what lets agent reputation be scored against real outcomes.

04Beyond the hackathon

The same pipeline, read instead of obeyed.

Strip out the scheduler and the auto-execution step, and what’s left is a research terminal: a per-ticker feed of AI conviction versus market pricing, a reviewed reasoning trail, and a risk gate — the same shape as the pre-trade workflow already built into a desk’s existing terminal, just running on an open pipeline instead of a closed one.

Point the graph at a scan a person triggers instead of a cron schedule, and it becomes something an analyst consults before advising a client — not a bot trading a book unattended. The autonomous loop demonstrated for the hackathon is one deployment mode of this system, not the only one it’s built for.

05What's next

Where this goes from a hackathon build.

shipped

Single-leg options, one name at a time

The Options Architect currently prices the nearest-ATM contract as a single long option — the simplest strategy that could prove the pipeline end to end.

next

Defined-risk spreads

Vertical and calendar spreads, chosen by comparing several constructions' risk/reward against the diligence-reviewed edge, instead of always reaching for an unhedged long option.

next

Portfolio-level Greeks

Today's risk gate looks at one trade and a position count. The natural next step is aggregate delta/vega exposure and correlation across open positions.

later

A human-in-the-loop terminal mode

Scans triggered by an analyst, recommendations surfaced for sign-off, execution requiring a click — the same reasoning trail, a different consumer.

later

A live-mode readiness gate

Not a toggle. A staged path with its own audit trail, position-size ramps, and kill-switch — treated with the same suspicion this system already applies to its own trade ideas.

Paper trading only · no live orders are ever sent
LangGraph orchestration·Alpaca market data & execution·33 backend tests