Towards Data Science

216 articles in the TruthFoundry index. Each links out to the original.

• Disaggregated GPU Inference Only Pays Off at Massive Scale; Chunked Prefill Better for Most Teams · Most teams should use chunked prefill instead of disaggregation, which only pays off with roughly 1,000+ GPUs and favorable networks.
• Power BI Developers: Five Microsoft Fabric Changes That Matter, What to Ignore · Microsoft Fabric changes Power BI licensing, storage, and Copilot while keeping core skills intact; Spark and KQL are safe to ignore for most developers.
• Run Multiple Claude Code Agents on a Cheap Rented CPU Server · Renting a remote CPU server via SSH lets you run 10-20 coding agents like Claude Code without buying powerful hardware.
• Data Analyst Turns Churn Prediction Model into Usable FastAPI Service · Data analyst builds churn model, then exposes it as a FastAPI service for real-world use.
• How to Preserve Table Structure in Enterprise RAG Systems · The article explains why flattening PDF tables breaks RAG and proposes four structural levels to handle them.
• Developer Builds Python Graph to Track Prompt Dependencies · A developer created a Python tool to map prompt dependencies and reduce retesting scope.
• Framework to Reduce Uncertainty in Agentic AI Development · A six-step framework helps teams define problems clearly before deploying AI agents to prevent costly rework.
• Fuzzy Matching Pitfalls: Damerau-Levenshtein Merges Distinct Sensor IDs · Comparing five string metrics shows no threshold separates typos from distinct sensor part numbers in a data lake.
• Enterprise RAG 'Not in Document' Answers Need Evidence From Four Pipeline Bricks · This technical article argues that RAG systems must supply four types of evidence to justify 'not in this document' answers.
• Graph Neural Networks Explained: GCN, MPNN, and GAT Concepts · The article explains how Graph Neural Networks use convolutions to process graph-structured data like molecules and social networks.
• PySpark Window Functions: Ranking, Totals, and Moving Averages · A guide to using PySpark window functions for ranking, running totals, and moving averages without collapsing data rows.
• Schema-Valid JSON Often Contains Hallucinated or Contradictory Data · Constrained decoding produces structurally valid JSON that frequently contains semantic errors like enum hallucination and cross-field contradictions.
• Why Missing Data Values Reveal Critical Insights About Processes · Statistical analysis shows missing data often encodes information about the underlying process rather than just being a nuisance.
• Bayesian Neural Networks Quantify Uncertainty Beyond Point Predictions · Bayesian Neural Networks replace single point estimates with probability distributions to quantify model uncertainty.
• 5 AI Skills Data Scientists Need to Stay Relevant in 2027 · Sara Nóbrega outlines five production AI skills—RAG, routing, guardrails, evals, and agent loops—that will keep data scientists relevant in 2027.
• Developer warns Structured Outputs guarantee shape, not truth · A developer explains that LLM Structured Outputs ensure valid JSON but can still hallucinate values if schemas are not designed for uncertainty.
• Designing RAG Systems by Authoring Your Own FAQ Corpus · A structured FAQ corpus simplifies RAG architecture by turning retrieval into a cache lookup and reducing generation costs.
• RAG Complexity Should Be Earned by Addressing Retrieval Failures · Experts argue RAG system complexity must address specific retrieval failures rather than being adopted as a default architectural pattern.
• AgentOps Monitoring Stacks Fail When AI Agents Loop in Production · Traditional monitoring assumptions break when AI agents execute multi-step loops, causing false green traces.
• 8 Tips for Writing Effective AI Agent Instructions · An expert outlines eight strategies for creating clear, maintainable, and robust instructions for autonomous AI agents.
• Anthropic Updates Context Engineering Rules for Data Scientists · Anthropic releases new guidelines shifting focus from prompt engineering to context engineering for Claude models.
• Classical Spell-Check Fails on Enterprise RAG Noisy Text · Traditional spell-correction tools struggle with enterprise RAG due to valid-word typos and boundary errors.
• 4 Advanced Claude Skills for Data Scientists in 2026 · The author outlines four advanced Claude skills for data scientists, including deep research, HTML briefs, design, and code documentation.
• Developer Details When to Use Claude Code vs Codex AI Coding Agents · A developer shares his rule: use Codex for single difficult tasks and Claude Code for orchestrating many smaller coding jobs.
• Bonus Series on RAG Details Six Cheaper NLP Techniques and Cross-Cutting Concerns · The article introduces a bonus series of six practical NLP techniques that complement RAG pipelines, covering spelling, OCR noise, tables, and local models.
• AI Agent Safety: Replace Broad Human Approval with Risk-Based Routing · A text-to-SQL agent caused near-disaster via broad human approval queues, solved by risk-based routing.
• How to Deliberately Use Codex Subagents for Complex Tasks · The article explains how to define and deploy specialist subagents within the Codex system to handle complex workflows.
• Developer Tests LangGraph AI Agent Postgres Backend with Docker · A developer demonstrates connecting a LangGraph booking agent to Postgres using Docker and hosted cloud databases.
• Why Claude Code Time Estimates Are Poor and How to Fix Them · The author argues Claude Code overestimates task duration due to human-biased training data and offers two methods to improve accuracy.
• The Sigmoid Function: Origins of 'e' and Neural Network Activation · The article explains the mathematical derivation of the sigmoid function using compound interest and calculus.