Index / Notes
Notes from ixprt.
Writing on data-for-AI, vector retrieval, quant infrastructure, and the work between raw inputs and useful AI.
-
Can AI Write Equity Research? The 2026 State of the Art
AI can draft the mechanical layers of equity research in 2026: earnings recaps, filing summaries, and estimate updates grounded in ingested documents. Thesis origination, accountability, and judgment calls remain human work.
Read → -
Kalshi vs Polymarket: How the Two Prediction Markets Actually Compare in 2026
A 2026 side-by-side of Kalshi and Polymarket: regulatory standing, liquidity by category, fee structures, contract design, and where each one is brittle.
Read → -
Positive Expected Value Betting, Explained for 2026
Positive expected value (+EV) betting is placing wagers only when the offered payout exceeds what the outcome's true probability justifies. The edge is computed as EV = p * profit - (1 - p) * stake, with the true probability usually derived by devigging a sharp book's price or a sharp-market consensus.
Read → -
RAG Chunking Strategies in 2026: A Practical Field Guide
A chunking strategy is the rule that divides documents into the units a RAG system embeds and retrieves. The 2026 field narrows to six practical approaches: fixed-size windows, structure-aware recursive splitting, semantic splits, late chunking, parent-document retrieval, and LLM-derived propositions, chosen by document type and budget.
Read → -
How to Remove the Vig From Betting Odds: A 2026 Guide
Removing the vig means converting a market's posted odds into implied probabilities, then rescaling those probabilities so they sum to exactly 100%; the rescaled numbers are the fair probabilities the bookmaker's margin was hiding. Four standard techniques do the rescaling: multiplicative, additive, Shin's method, and the power method.
Read → -
Reranking in RAG Pipelines: When Cross-Encoders Pay Off
Reranking adds a second scoring pass to a RAG pipeline: a cross-encoder rescores the retriever's top candidates so the best chunks reach the model's context window. It earns its latency once the correct document reliably lands in the top 50 results but outside the top 5.
Read → -
How Diagest Powers CLV.gg's Three-Year Historical Betting Database
CLV.gg's live edge detection sits on top of three-plus years of historical betting prices, billions of records of lines, closes, and settlements. That history is only useful because Diagest, the ixprt data-for-AI pipeline, cleans, aligns, and serves it. Here is how the two systems work together and what the history makes possible.
Read → -
What Is CLV.gg? Inside the Real-Time Betting Intelligence Platform
CLV.gg is the live sports-betting intelligence platform built at ixprt. It reads the covered market across a dozen-plus books and venues in real time, builds a sharp-weighted estimate of the fair price, and surfaces the edges while they are still takeable. Here is what it is, why the speed matters, what it finds, and how deep the book coverage runs.
Read → -
AI Agent Memory Architectures: How Persistent Memory Actually Works
The four patterns for giving AI agents persistent memory in 2026, the specific systems shipping each one (Letta, Mem0, OpenAI, Claude), and the recall-fidelity vs token-cost tradeoff every team eventually confronts.
Read → -
How to Evaluate RAG Pipelines
A 2026 buyer's guide to RAG evaluation: the three eval families, the metrics that actually predict production quality, the four major frameworks (Ragas, TruLens, ARES, DeepEval) compared, and which one to start with depending on where the pipeline is in its lifecycle.
Read → -
What Is an AI Analyst Data Pipeline? 2026 Field Guide
What an AI analyst data pipeline is, why it's the load-bearing layer beneath every credible analyst agent, and how production systems are being built in 2026.
Read → -
Closing Line Value in 2026: A Field Guide
What closing line value (CLV) is, why it's the only public-data-verifiable proxy for betting edge, and how the 2026 generation of CLV platforms is reshaping serious sports betting.
Read → -
Per-Corpus Schedule Modes: Why Ingestion Settings Belong on the Data Source, Not the Pipeline
A pipeline-wide ingestion mode is a category error. Production data-for-AI systems need a per-corpus schedule mode because each source has a different cost, cadence, and failure profile.
Read → -
How to Build a Claude Code Operating Agent: The 2026 Starter Template Guide
A 2026 buyer's guide to Claude Code starter templates. What a production-ready agent template includes (hooks, memory, skills, auto-commit, guard rails), how to set one up in 5 minutes, and how to pick between building from scratch and forking claude-agent-starter.
Read → -
Honest Zeros: Why a New Data-Pipeline Dashboard Should Launch With Empty Cells
Most data-pipeline dashboards launch with synthetic baselines that look healthy on day one. The teams that ship retrieval at scale launch with honest zeros and let the numbers earn their position.
Read → -
Why Retrieval Drift Goes Undetected: The Case for Public Pipeline Dashboards
Most teams discover RAG pipeline drift through customer complaints. The teams that catch it first publish their drift, recall, and ingestion-run health to a page anyone can read.
Read → -
Scheduled vs. Event-Driven Ingestion: A 2026 Comparison for Data-for-AI Pipelines
Scheduled ingestion and event-driven ingestion are framed as opposites. In production, they are two halves of the same operating model, and most teams pick the wrong one first.
Read → -
Embedding Model Migrations Without Re-Indexing Everything: A 2026 Playbook
A 2026 playbook for migrating between embedding models (OpenAI, Voyage, Cohere, open-weight) without re-indexing the whole corpus the first weekend. Covers the four migration patterns that work in production, the costs each one pays, and the operating discipline that keeps retrieval honest during the switch.
Read → -
MCP vs a Custom HTTP API: When Each One Is the Right Call in 2026
A 2026 decision framework for choosing between a Model Context Protocol server and a plain HTTP API for any service that AI agents will call. Covers the four signals that decide the call, the cases where MCP is the wrong answer, and the migration path that lets a team ship both without doubling the engineering bill.
Read → -
Reading the Fed in 2026: RSS, Blob Fetching, and What You Actually Get From Public Federal Reserve Data
A 2026 walkthrough of what the Federal Reserve actually publishes on RSS, where the bodies live (almost always behind another link), and how to build an ingestion pipeline that does not lose the document in the index.
Read → -
Backtesting AI Agent Signals: Pitfalls and Patterns
Backtesting AI-generated signals is not backtesting strategies. The input shape differs, the failure modes differ, and the standard quant tooling assumes things that don't hold. A field guide to what breaks and what works.
Read → -
AI Analyst Voice Consistency: Why Personas Drift and What to Do About It
Multi-month AI analyst products have a quiet failure mode: the persona stops sounding like itself. Why it happens and how the field handles it in 2026.
Read → -
Document Parsing for AI: a 2026 Strategy Guide
How modern document parsers compare in 2026: layout-aware, naïve, and LLM-based approaches stacked across PDF, DOCX, HTML, and scanned formats.
Read → -
Risk Attribution Models Compared: Barra vs Axioma vs Custom Approaches
How the major risk-attribution approaches compare: Barra, Axioma, and the custom factor models that funds increasingly build in-house.
Read → -
Vector Store Choices in 2026: Qdrant vs Pinecone vs pgvector vs Weaviate vs Milvus
Five vector stores side by side: feature matrix, pricing posture, latency at scale, and which to pick by use case.
Read → -
AI Analyst Desks: A 2026 Field Guide
What AI analyst desks are, what makes them work, and where they fit in the 2026 market-research landscape.
Read → -
What is a Quant Engine? A Buyer's Guide for Funds and Family Offices
A definition of the quant-engine category, the four functions it covers, and what funds and family offices should evaluate before buying or building.
Read → -
Why RAG Pipelines Fail: 5 Common Pitfalls (and What to Watch For)
Five named failure modes that kill RAG systems in production: drift, dedup gaps, chunk-strategy mistakes, retrieval-recall miss, and embedding-model mismatch.
Read → -
What is Data-for-AI? A Buyer's Guide to the Modern Stack
A definition of the data-for-AI category, the five layers of work it covers, and what to look for when you evaluate vendors.
Read → -
The State of AI in Finance: 2026 Benchmark
Where AI sits in finance in 2026 — data infrastructure adoption, quant integration, AI-generated equity research market, regulatory framing, and a 12-month outlook.
Read →