Research notes
Notes on AI in finance.
Where the agentic moment meets capital markets: what is happening, why it matters, and what I am building toward.
RED QUEEN: Training an AML Detector Against Crimes That Do Not Exist Yet
Conventional transaction surveillance can only catch the patterns someone already wrote a rule for. It is structurally blind to the typology nobody has seen. RED QUEEN manufactures those unknown-unknowns on purpose, in a closed synthetic loop, and hardens a detector against them before they appear in the wild.
DELPHI: Build the Skeptic Into the Architecture, Not the Prompt
LLMs fail at finance in one specific way: a confident model recalls a plausible but wrong figure, and a single-pass pipeline has nothing to catch it. Sell-side research solved this structurally decades ago — specialists draft, a skeptic challenges, compliance gates publication. DELPHI makes that workflow the architecture.
CADUCEUS: Pressure-Testing Agentic AI on the Highest-Stakes Domain There Is
A molecular tumor board is the textbook case for multi-agent AI — and the textbook case for getting AI in medicine wrong. CADUCEUS convenes 55+ specialist agents across 7 layers, argues against its own consensus, and lets nothing reach a human without a retrievable source. The patterns are the finance patterns, on the hardest possible stakes.
PRAETOR: In M&A Diligence, the Value Is in the Connections — So I Made an Agent That Cross-References
M&A due diligence is thousands of documents, a hard exclusivity clock, and value that lives in the threads across workstreams — a lawsuit Legal finds is a provision Finance must book. Humans lose those threads across siloed teams. PRAETOR runs 9 specialist agents in parallel, cross-references their findings into one register, and renders a cited verdict.
AEGIS Live: Real-Time AML Where Latency Is a Feature and Train/Serve Skew Is the Enemy
Money laundering is confirmed retrospectively — live blockchain data never arrives with a "laundered: yes/no" label. So you score in real time and humans confirm later. AEGIS is the full streaming stack: resilient ingestion, a backpressure bus, a sliding-window graph, a graph/ML ensemble, single-digit-ms p95, and a parity gate that kills train/serve skew.
AEOLUS: Everyone Predicts Failures — The Money Is in Deciding When to Act
Failure prediction is commoditised. The hard, valuable problem in industrial AI is the closed loop after the prediction: deciding when to act so you lose the least revenue, under real constraints, with a governance trail an operator can deploy. Here is how I built that for a wind fleet.
QUORUM: An AI Investment Committee That Argues Before It Decides — and Never Makes Up a Number
Most "AI investing" demos fail the same way: a confident model recalls a wrong figure. QUORUM splits the system in two — Python computes every number deterministically, the LLM only argues and narrates. Six agents debate from real market data and converge on a documented allocation.
RegRadar: An Agentic Engine That Reads EU Regulation Without Hallucinating a Single Citation
Every act in the EU regulatory firehose carries concrete obligations a bank must implement. RegRadar extracts them, maps them to systems, ranks them by deadline, and drafts the gap-assessment memo — with every claim verified, programmatically, against the live EUR-Lex source. Here is the architecture that makes it trustworthy.
CreditForge: What It Actually Takes to Build a Credit-Risk Model a Bank Would Sign Off On
EL = PD × LGD × EAD is one line. The 80% that makes a credit model bank-credible is everything around it — leakage-safe point-in-time targets, out-of-time validation, calibration, reason codes, fairness, and drift monitoring. Here is the full stack.
Building Recoupe — Why Subrogation Is the Best Vertical to Pressure-Test Agentic AI
A multi-agent system that reads insurance claims, assigns fault by jurisdiction, computes what is recoverable, and pursues it — with grounded citations and an append-only audit trail. Here is what I learned shipping it.
JPMorgan's $2B AI Bet: What 2,000 Use Cases Tell Us About Banking's Next Decade
Jamie Dimon spent $2B on AI in 2024 and built 2,000+ use cases. Here's what that signals for every engineer wanting to work in banking.
The Bloomberg Terminal Is About to Be Eaten by an Agent
The $32K/year Terminal is the most valuable software seat in finance. AI agents can replicate 70% of its workflows for $99/month. Here is how the disruption plays out.
Why Goldman's IndexGPT Is the First Real Glimpse of Agentic Wealth Management
Goldman patented IndexGPT in 2024 — an AI system that constructs custom investment indices on demand. It is a tiny product. It is also the future of wealth management.
The Quant ML Revolution: How Two Sigma and Renaissance Are Rebuilding for the LLM Era
Traditional factor investing is dying. The funds that survive the next decade will be the ones that figured out how to ingest unstructured data at scale. Here is the shift in progress.
MCP and the Quiet Death of Fintech APIs as We Know Them
Anthropic's Model Context Protocol launched in late 2024. Most people read it as a developer convenience. Inside fintech, it is the start of a much bigger rewrite.