Applied intelligence · Finance · Systems
Where intelligent systems meet financial decisions.
Explore the ideas, architectures, and experiments connecting AI agents, quantitative finance, and digital assets, from first principles to working concepts.
Explore the Lab Read the Research
Research and concept demonstrations. Not investment advice.
The intelligence stack
Five stages take raw data to a monitored decision. Select a stage to see what it takes in, what it hands on, and the guardrail that keeps it honest.
01 Data intake
Public or synthetic inputs are collected, timestamped and checked for gaps before anything downstream sees them. Bad data stopped here is the cheapest error the system will ever catch.
- Takes in
- Prices, filings, on-chain records
- Hands on
- Clean, versioned datasets
- Guardrail
- Schema checks; gaps and outliers flagged
02 Models and agents
Statistical models and bounded AI agents turn clean inputs into estimates. Each estimate carries a stated confidence and a list of conditions under which it is known to fail.
- Takes in
- Validated features
- Hands on
- Estimates with uncertainty
- Guardrail
- Agents act only within written permissions
03 Decision support
Outputs become explanations a person can challenge: what changed, why, and how sure the system is. The system recommends; a named human decides.
- Takes in
- Estimates and their rationale
- Hands on
- A recommendation for review
- Guardrail
- No action without human approval
04 Execution sandbox
Approved ideas run against simulated markets only. Nothing in this layer touches a brokerage account, a wallet or a real order book.
- Takes in
- An approved scenario
- Hands on
- Simulated fills and profit and loss
- Guardrail
- Simulation only; no live connectivity
05 Monitoring
Drift, drawdown and data-quality alarms watch the whole loop and send problems back to the start, where they can be fixed rather than explained away.
- Takes in
- Sandbox and model metrics
- Hands on
- Alerts, rollbacks and reviews
- Guardrail
- Every alert is logged and reviewed
Working concepts
Each concept is a working instrument. Change an input and the output changes for a reason you can read.
-
Simulation
Quant Finance Lab
Generate a synthetic market, run a simple trend rule against it and watch return, volatility and drawdown respond.
-
Concept
Agent pipeline with a human in the loop
Follow three sample cases through validation, analysis and explanation to a person who approves or rejects them.
-
Concept
Tokenization lifecycle
Step through how a right becomes a token, who holds it, and what must be true for it to be redeemed.
-
Simulation
Financial system digital twin
Apply liquidity, concentration and price shocks to five linked institutions and see where stress travels.
Latest research
All research-
Read the drawdown before the return
A headline return hides the path taken to reach it. Maximum drawdown and time under water describe what holding the strategy would actually have felt like.
Quant financeSample content
-
Why a beautiful backtest should worry you
Try enough settings on one history and something will look brilliant by luck. Demo notes on overfitting, using the Quant Lab’s random seed as the control.
Quant financeAI & agentsSample content
-
Agents on a leash: bounded autonomy in financial workflows
Research note on designing AI agents for finance around permissions, confidence floors and human approval, mapped to the NIST AI RMF and the EU AI Act.
AI & agentsFinancial systemsSample content
How to read this site
Sources are primary and dated
Factual claims link to the regulator, standards body or central bank that published them, with the date we checked the page.
Simulations say so
Every chart built on generated data is labelled synthetic, uses a fixed seed and gives the same result every time.
Limits are stated next to results
Models here are deliberately simple teaching tools. Their assumptions and blind spots sit beside their outputs, not in a footnote.