miramarlabs.ai
Agentic AI · Public Markets · Human-in-the-Loop

An agentic AI investment platform.

Event-driven signal generation. Bounded LLM reasoning over raw market data. Independent rule-based validation with an automatic retry cycle. A non-bypassable human approval gateway. No order reaches the brokerage without a human in the loop.

See how it works →
Accumulation, reasoning, catalyst — an abstract composition evoking the platform's intraday pattern engine and the chemical catalyst reaction that inspired its name.

Architecture pillars

Agentic research

A continuous pipeline of specialist reasoners — cross-market risk, institutional flow, sector rotation, momentum and reversal patterns, longer-horizon trend awareness. Each produces a structured artifact; a synthesis agent assembles a ranked briefing the primary reasoning agent inherits.

LLM reasoning engine

A bounded ReAct loop with a tool registry over raw market data — not pre-cooked composite scores. Technicals, fundamentals, options flow, institutional flow, reference commentary, reversal trajectory, relative strength, and intraday indicators are callable tools.

Human approval gateway

A non-bypassable node in the execution graph. Every trade proposal is routed to a human operator via Slack. No order reaches the brokerage API without an affirmative tap. Enforced by construction — there is no other execution path.

Intraday entry and exit timing

A separate agent decides when within the trading day to open — and close — an approved position. Pattern-based entry timing sits behind a universal risk gate; volatility-aware exit guardrails manage the close.

Full audit trail

Every decision, action, and outcome lands in an append-only ledger, linked so any fill traces back to the reasoning that produced it. Complete by construction — meeting the audit-trail requirements the EU AI Act in regulated industries.

Evaluation and reflection loops

Trajectories are graded per-fill by returns math and an LLM judge once outcomes settle; end-of-day and monthly post-mortems review the broader corpus. Aggregate verdicts feed a calibration layer injected into the next round of reasoning. Learning without gradient updates.

What's inside