Insights
Data Science insights
61 articles on data science — written by the engineers doing the work, published by OmniDataTec.
Articles
9 min read
AI Agents Plus Gurobi: Where the LLM Stops and the Solver Starts
Where to draw the boundary between a language model and a MILP solver in a production decision system, and what breaks when the boundary is fuzzy.
5 min read
Account-Takeover Detection With Session and Device Graphs
Account takeover is a change-of-actor problem, not an anomaly problem. How session and device graphs give you features that per-account modelling cannot.
7 min read
Agentic Decision Intelligence for Budget Allocation
What an agent adds to budget allocation that a scheduled solver job does not, where the autonomy dial should sit, and how the pattern fails in production.
7 min read
Alert-Triage Agents: Cutting AML Backlogs Without Cutting Coverage
How to use agents to clear transaction-monitoring backlogs without quietly narrowing detection coverage, and where the human decision must stay.
7 min read
Architecting an AI Agent for AML Alert Investigation
A reference decomposition for an alert-investigation agent: what it gathers, what it may decide, and the failure modes that only show up after deployment.
7 min read
Beneficial-Ownership Analysis With Knowledge Graphs
Ownership and control are two different graphs with different arithmetic. How to model them so a UBO answer survives review, and where the data breaks.
9 min read
Building a Walk-Forward Backtesting Harness That Does Not Lie to You
A backtest is a measurement instrument. What the harness must own — folds, point-in-time features, costs, trial counting — for the number to mean anything.
8 min read
Calibration Beats AUC: Scoring Financial Risk Models the Way Decisions Use Them
AUC measures ranking; financial decisions consume probabilities. How to measure and fix calibration on risk models that feed thresholds and expected loss.
7 min read
Churn Modelling in 2026: Where Gradient Boosting Still Beats Tabular Foundation Models
A practitioner's decision rule for when a tabular foundation model beats gradient boosting on a churn table — and when it cannot.
7 min read
Constrained Decision Making When the Business Rules Are Non-Negotiable
How to encode genuinely hard business rules in a decision system, why infeasibility is useful output, and how to govern a constraint set nobody fully owns.
7 min read
Cost per Successful Task: A Better Metric for AI Systems
Cost per attempt rewards cheap failure. Here is how to define success, compute cost per successful task honestly, and where the metric itself misleads.
5 min read
Crypto and Stablecoin Flows: What Changes in Transaction Monitoring
A measured look at what stablecoin and on-chain settlement actually change for transaction monitoring architecture — and what stays exactly the same.
5 min read
Customer Lifetime Value Models That Survive Contact With Finance
Most CLV models die in the first meeting with Finance. The four questions they always ask, and how to build a model that answers them.
7 min read
Designing Financial Crime Agents That Can Explain Every Action
Auditability in financial crime AI is action provenance, not model interpretability. How to design an agent whose every step is reconstructable months later.
10 min read
Entity Resolution Is the Prerequisite Nobody Budgets For in Financial Crime AI
Every graph, screening engine and investigation agent inherits its ceiling from the entity-resolution layer. How to build that layer, and how it fails.
7 min read
Evaluating Financial Models Across Market Regimes, Not Just Test Splits
A single test-set score averages over conditions the model will meet separately. How to define regimes without leaking, and what to report instead.
7 min read
Event-Driven Agents for Real-Time Transaction Monitoring
Where an LLM agent can and cannot sit in a payments path: latency tiers, a streaming topology, and the duplicate-action problem nobody budgets for.
8 min read
Feature Engineering Credit-Risk Signals From Raw Transaction Records
Which signals in a raw transaction ledger carry credit information, how to window them without leaking, and the joins that quietly invalidate a backtest.
6 min read
Feature Stores for Payments Data Science: What They Fix and What They Do Not
A feature store fixes training/serving skew and point-in-time joins. It does not fix definition ownership, entity identity or drift, and payments adds more.
7 min read
Fraud-Ring Detection With Graph ML and Community Structure
Modularity-based community detection finds the wrong communities on fraud graphs. What to run instead, and how to evaluate detection at ring level.
Page 1 of 4