Five layers between raw data
and executed trades.
Arkraft is an AI-native quant hedge fund stack. Each layer provides well-defined outputs to the layer above — like departments in a financial institution, but orchestrated by specialized AI agents.
Select a layer to explore its details
L0 → L4
From raw data to executed trade
Follow a single security through all five layers — from ID resolution to order fill.
NVDA → canonical ID: US67066G1040
Ingest NVDA earnings call transcript + price data
Compute sentiment scores + momentum features
Alpha agent generates BUY signal, portfolio agent approves
Order routed to NYSE, filled at $892.34, 0.3 bps slippage
How the layers work together
Layer isolation
Each layer exposes a clean interface. Consumers never reach past one layer. L3 talks to L2, never directly to L1. This means you can swap a data provider in L1 without touching anything in L3 or L4.
L3 agents query L2's feature store — they never parse raw CSV files from L1.
AI-native
L3 agents are first-class citizens, not bolted-on tools. They have persistent memory, learn from outcomes, and collaborate with each other. The intelligence layer isn't a wrapper around an API — it's the core of the system.
An alpha agent that discovers a signal also validates it, monitors it live, and detects when it decays.
Reproducibility
Every pipeline step is versioned and auditable. Given the same inputs and the same configuration version, you get the same outputs. This isn't just good engineering — it's a regulatory requirement.
Any trade can be traced back through the exact signal version, data snapshot, and agent reasoning that produced it.
Incremental build
Layers can be developed and deployed independently. Start with L0 + L1 for data infrastructure, add L2 for warehousing, then bring intelligence online. Each layer delivers value on its own.
A firm can run L0–L2 for a clean data platform, then add L3 agents one at a time as strategies mature.