Skip to main content
Research stage Research-stage project: no live trading, no investment products, and no performance results. Read the risk disclosures
MAA
Menu

Platform

Four layers, one research discipline.

An overview of the platform’s design goals. These describe what is being researched and built, not a live service.

Research

Market data ingestion

Research is only as good as its data. Ingestion is designed around provenance and point-in-time correctness.

  • Broker and vendor adapters behind MCP servers, unified by UBI, so each source is isolated and independently permissioned.
  • Normalized schemas with explicit as-of timestamps to prevent look-ahead leakage.
  • Provenance metadata on every record: source, retrieval time, and license terms.
  • Automated quality checks for gaps, duplicates, stale quotes, and unexpected revisions.
  • Handling of corporate actions and symbol changes as a first-class concern.
Research

Agent orchestration

A Python harness coordinates specialized agents and keeps every run bounded, validated, and logged.

  • Role-specialized agents with narrow responsibilities rather than one general-purpose prompt.
  • Structured outputs validated against schemas; malformed output is rejected, not repaired silently.
  • Tool access mediated through MCP servers with least privilege.
  • Run budgets for time, tokens, and tool calls, with deterministic seeds where supported.
  • Complete recording of prompts, responses, tool calls, and model versions.
Research

Research workflows

Experiments follow a written process so conclusions are not shaped by the results they produce.

  • Hypothesis, data specification, experiment, and review as explicit, recorded steps.
  • Versioned configurations and an experiment registry, including negative results.
  • Out-of-sample and walk-forward validation, with attention to multiple-testing effects.
  • Cost, slippage, and liquidity assumptions stated up front.
  • All historical simulations labeled as simulations, with their limitations.
Research

Risk controls

Controls are enforced by deterministic code that models cannot talk their way around.

  • A policy engine that evaluates proposed actions against rules expressed as code.
  • Exposure and concentration limits, plus hard stops, defined outside any prompt.
  • Human approval required for anything beyond research output.
  • A kill switch and a complete, immutable decision log.
  • No live trading. Any order capability is limited to simulated (paper) accounts, off by default, and gated by the policy engine.

Design goals can change as research progresses. Statuses reflect the current stage, not a delivery commitment.