Insights
Machine Learning insights
46 articles on machine learning — written by the engineers doing the work, published by OmniDataTec.
Articles
8 min read
Are Tabular Foundation Models Finally Challenging XGBoost?
TabPFN-2.5 reports beating default XGBoost on every small table. Read the word default carefully, then decide what to move.
5 min read
AutoML in the Foundation-Model Era: What PyCaret-Style Tools Are Still For
Tabular foundation models changed what a baseline costs. The case for keeping a low-code AutoML layer, and the three jobs it still does better than anything.
7 min read
Automated Feature Engineering With Coding Agents: What to Delegate
Coding agents write feature code faster than any team can review it. A gate-based delegation model for deciding what they may generate and what they may not.
6 min read
Benchmark Leakage in Time-Series Foundation Model Evaluations
Pre-training overlap inflates reported forecasting accuracy by 47-184%. How the contamination happens, and how to evaluate so it cannot.
6 min read
Building Graph-Native AI Agents
What changes when an agent's working state, evidence and provenance live in the graph rather than in its context window — and the costs that arrive with it.
7 min read
Building a Customer Knowledge Graph From Fragmented Enterprise Records
A delivery guide for turning CRM, billing, support and third-party records into one customer graph — hierarchy, survivorship, incremental loads and stewardship.
6 min read
Calibration of Probabilistic Foundation Forecasts
A zero-shot model's 95% interval is a claim about your data it has no way to make. How to test it, and how to fix it without retraining.
8 min read
Can Zero-Shot Forecasting Replace a Custom Treasury Model?
The honest answer is no, and the reasons are not about accuracy. What zero-shot forecasting does replace, and a staged plan that does not risk the plan.
9 min read
Chronos-2 vs. TimesFM vs. Moirai for Financial Forecasting
Three forecasting foundation models, three different architectural bets. What the leaderboards say, what leakage takes back, and how to choose.
6 min read
Combining Graph Algorithms With LLM Reasoning Without Letting the LLM Traverse
An architecture where the LLM chooses and explains but the graph engine walks: bounded traversal tools, a Cypher validator, and the failures each one prevents.
6 min read
Distilling a Tabular Foundation Model Into Something You Can Actually Serve
In-context models carry their training set to every inference. Distillation removes that. Here is the transfer set, the pipeline and what it costs.
5 min read
Embeddings for Structured Data: What Works Outside NLP
Where learned representations beat one-hot and target encoding on tabular and relational data, which patterns hold up in production, and which quietly fail.
8 min read
Entity Resolution Quality Determines GraphRAG Answer Quality
How a false split or a false merge becomes a confidently wrong GraphRAG answer, and how to measure the coupling instead of assuming it away.
9 min read
Entity Resolution to Knowledge Graph to GraphRAG to Agent: A Reference Architecture
The four-layer stack behind a working GraphRAG system, what each layer owes the next, and which parts you can defer without poisoning everything downstream.
6 min read
Evaluating GraphRAG: Metrics Beyond Answer Similarity
Answer similarity hides the failures that matter in graph retrieval. Evaluate five layers: resolution, planning, retrieval, grounding and completeness.
8 min read
Foundation Models vs. LightGBM for Transaction-Volume Forecasting
Zero-shot time-series foundation models against a tuned LightGBM baseline for payment-volume forecasting, and how to run that comparison honestly.
5 min read
Graph Algorithms Every Financial Data Team Should Actually Use
Five graph algorithms that repay the engineering cost in financial data work, the failure modes each one hides, and the ones you can safely skip.
6 min read
Graph Data Quality: Bad Edges Are Worse Than Missing Ones
Why a false edge corrupts a graph non-locally while a missing edge only costs you one answer, and the write-time controls that follow from that asymmetry.
6 min read
Graph Embeddings vs. Handcrafted Graph Features: A 2026 Assessment
Where learned graph embeddings beat handcrafted structural features, where they do not, and the operational problems that appear after the second retraining.
5 min read
Graph Neural Networks for Financial Risk: Where They Earn Their Complexity
Gradient boosting on graph-derived features is a strong baseline. The narrow set of conditions under which message passing actually beats it.
Page 1 of 3