Ideas and field guides for reliable, reviewable AI-assisted data analysis.
A document diff does not explain why an interpretation changed. Change Context ties formal asset changes to their stated reasons and evidence references, while preserving unknowns and confirmation status.
Read article →Python made analysis flexible, but rapid agent edits made scripts, dependencies, and intermediate outputs harder to follow. Why I chose SQL-first, and where Python still belongs.
Read article →AI can deliver results faster than people can absorb them. Here is how I preserve decisions and evidence so the person responsible can explain, challenge, and own the analysis.
Read article →How I let external Agents keep their general capabilities, use a built-in Agent for a deeper analytics workflow, and connect both to one control plane.
Read article →Once AI Agents find data, write SQL, run notebooks, and revise reports, chat history and file versions no longer preserve the full evidence chain. This is why data analysis needs Analysis Lineage.
Read article →A Skill can preserve an analytical method and package SQL or context from a specific case. Real analysis also depends on data versions, execution state, human decisions, evidence, and limitations.
Read article →An AI agent does more than produce longer answers: it uses tools based on state, advances multi-step work, and leaves verifiable deliverables. This article explains the differences, evaluation criteria, use cases, and adoption boundaries through data analysis.
Read article →The hardest AI analysis errors to detect come from definitions, denominators, joins, and time windows. This article presents four review layers: Plan, SQL, checkpoints, and limitations.
Read article →Natural language lowers the barrier to querying data, but it cannot replace SQL as reviewable evidence for grain, joins, filters, and aggregation. The better division of work is AI drafting, SQL preserving evidence, and people reviewing it.
Read article →A Plan aligns the decision question, denominator, grain, time window, and validation approach before data retrieval. Execute then becomes a clear human authorization boundary, preventing ambiguous assumptions from turning into official numbers.
Read article →Learn how reproducible AI analysis preserves inputs, SQL, assumptions, validation, and reports so teams can rerun results and explain every change.
Read article →Analysis agents, conversational BI, and traditional BI serve different stages. Quick answers, definition exploration, and stable monitoring should not be confused. This article provides a five-dimensional selection matrix and three practical scenarios.
Read article →Learn how local-first AI data analysis reduces data movement, queries local files, and preserves reusable work while clarifying cloud model boundaries.
Read article →I want analysis context to narrow at decision points, not only at capacity limits. Preserve accepted definitions, unresolved questions, and evidence locations without turning tentative work into verified conclusions.
Read article →A Skill can record the last analysis correctly and still apply it in the wrong project. How I separate reusable methods, shared rules, project knowledge, and decisions for one analysis.
Read article →Prompt engineering is not obsolete, but agents must also manage tools, data, state, history, and memory. Context engineering determines what the model actually sees on each turn.
Read article →A larger context window can hold more tokens, but that does not mean a model uses every passage equally reliably. Long tasks need compaction, structured artifacts, and on-demand loading to control context rot.
Read article →Long-term agent memory affects analysis across conversations, so enterprise knowledge needs provenance, scope, approval, and retirement. This article provides a practical Memory governance framework and Queued Knowledge approval flow.
Read article →MCP connects AI applications, data, and tools through a standardized client-server protocol. This article explains its core components, security boundaries, and Lantide's two-way MCP integration.
Read article →MCP standardizes connections but does not guarantee least privilege. Use host, credential, scope, capability, approval, and audit controls to review agent data access.
Read article →AI Agent governance starts by defining capability boundaries among users, agents, tools, and data, then adding approval, monitoring, and recovery. Use this six-part capability map to assign owners and create testable controls.
Read article →Table names and field types describe the shape of data, not the definitions of valid customers, revenue, conversion, or time windows. Learn which business semantics AI analysis needs and how to turn definitions into a reviewable contract.
Read article →Use Lantide as a local MCP workspace for Claude Code and Codex, keeping plans, SQL, evidence, and reports reusable across data analysis sessions.
Read article →PostHog excels at product events, funnels, and experiment significance. When local CRM or finance data should not move to the cloud, or you need reviewable SQL and a formal Report, use Lantide for local diagnosis.
Read article →Hex excels at collaborative cloud notebooks, reactive execution, and app publishing. When data must stay local or ad hoc exploration does not deserve a cloud project, use Lantide for local-first analysis and send sanitized results back to Hex.
Read article →dbt is built for production transformations, tests, and CI. For one-off hypothesis validation, use Lantide as a local sandbox, then write standard SQL back to the dbt repository once the definition is proven.
Read article →Notion is suited to PRDs, knowledge bases, and database charts, but it is not a local SQL engine. Run rerunnable analysis in Lantide, then put the HTML or Markdown result back in Notion so planning pages cite evidence-backed conclusions.
Read article →Obsidian and Dataview excel at local Markdown knowledge and note metadata queries. Lantide adds DuckDB SQL, rerunnable analysis lineage, and governed Agent memory; the two connect through files rather than replacing each other.
Read article →A practical workflow for rerunning weekly CSV or Excel analysis while preserving inputs, metric definitions, SQL, validation checks, and report limitations.
Read article →AI can explore CSV files quickly, but it cannot define correct data for you. This checklist covers encoding, types, nulls, duplicates, keys, time zones, units, and outliers with rerunnable SQL checkpoints.
Read article →DuckDB is an embedded analytical database that queries local CSV, Parquet, and Excel files directly. It excels at file-based ad hoc analysis but does not replace a shared warehouse or scheduled ETL.
Read article →Successful execution does not mean AI-generated SQL answered the right question. Review grain, JOINs, filters, NULLs, time, denominators, and LIMITs with these minimal validation queries.
Read article →A funnel conversion rate has no single universal formula. User versus event grain, cohort versus calendar windows, event order, and cancellation rules all change the answer.
Read article →When totals grow after a JOIN, the aggregation is usually not broken—the tables have incompatible grains. Use these order examples and SQL checks to catch duplicate counting before delivery.
Read article →Quick and formal analysis differ not by question length but by the impact of a wrong answer and the chance it will be reused. This two-dimensional matrix helps choose the right workflow.
Read article →Data lineage traces sources, transformations, and outputs. Even with local CSV files and a monthly report, reusable work benefits from minimal lineage, naming rules, and a rerun path.
Read article →A trustworthy AI report needs more than a summary: decision question, definitions, findings, traceable evidence, limitations, and next steps. Use this six-part template tied to SQL evidence.
Read article →Data Context, curated domain knowledge, and past analytical experience serve different purposes. An event participation example shows what teams need to preserve for a useful handoff.
Read article →Access to data does not preserve the reasoning behind past work. Teams also need methods, validation, and decision context so another analyst or agent can continue the analysis.
Read article →A PM need not rewrite SQL line by line. Confirm the decision, population, denominator, time, exclusions, and validation before Execute to protect business definitions and accountability.
Read article →Agents can accelerate schema exploration, SQL drafts, and report drafts. Analysts should own problem definition, business metrics, validation, anomaly interpretation, and decision communication.
Read article →Before connecting an external Agent, define sources, capabilities, workspace boundaries, credential lifecycle, formal evidence, writer ownership, audit, and recovery.
Read article →A 30-day AI analysis pilot should begin with a real, controlled retrospective, then establish quality, review failures, permissions, and expansion criteria.
Read article →