Human approval for AI agents: review the exact action
Design approval requests with clear payloads, reviewers, expiry and safe recovery instead of a vague approve button.
Read moreTHE ENGINEERING PUBLICATION
Engineering Notes
Technical guides, experiment methods, architecture breakdowns and things learned by building.
Twenty guides, with tools and examples to try.
Start with the Mac benchmark results, agent reliability checks, or a workflow you want to automate.
Inspect the research and evidence →Explore the work
Agent retries and idempotency: recover without duplicate writes
Design retry budgets, logical action identifiers and reconciliation for agents calling external systems.
Read moreAI agents for SEO automation: build an evidence-to-review pipeline
Design an SEO automation workflow that collects crawl evidence, drafts fixes, validates output and measures changes without invented ranking gains.
Read moreAI automation consulting pricing: scope the cost you can verify
Understand discovery, implementation and operating costs with transparent worked examples and a practical automation scoping checklist.
Read moreAI workflow automation tools: choose the system around the job
Compare n8n, Flowise and custom code using workflow ownership, approval boundaries, deployment needs and a practical acceptance test.
Read moreCSV to JSON for automation: preserve data before connecting tools
Handle quoted fields, leading zeros, duplicate headers and explicit types when converting CSV datasets into JSON workflows.
Read moreHow to evaluate AI agent accuracy and reliability
Build outcome-based agent tests, separate safety from task success, and understand pass-rate uncertainty with a worked dataset.
Read moreM2 Max 64GB: 81 tok/s solo, 213 tok/s with 4× batching
Our M2 Max 64GB oMLX capture shows 81.0 tok/s solo and 213.1 tok/s aggregate at 4× batching, while E2E latency rises from 2.662 to 8.526 seconds.
Read moreLLM context, KV cache and memory bandwidth: why RAM is only one constraint
Understand weight capacity, KV-cache growth and Apple silicon bandwidth without turning hardware specifications into token-speed promises.
Read moreRun a local LLM on Mac: a workload-first setup guide
Plan local AI on Apple silicon, choose a starting interface, budget memory and verify that your intended workload remains local.
Read moreLocal versus cloud AI costs: compare useful completions
Compare capital cost, API usage, operator time and task quality with transparent assumptions rather than a universal cheaper option.
Read moreMLX vs Ollama on Mac: compare the workflow before the speed
A documented MLX LM and Ollama comparison covering model artifacts, integration needs, memory planning and a fair benchmark protocol.
Read moreHow to monitor AI agents in production: trace the outcome
Design agent traces, meaningful service metrics, failure alerts and a recovery runbook without confusing model output with completed work.
Read moren8n in production: write the failure runbook before adding workflows
Prepare ownership, credentials, execution recovery and acceptance tests for an operational n8n deployment.
Read moreDefine the boundaries before building an AI agent
A design note on permissions, approval gates and failure recovery for tool-using agents.
Read moreRAG evaluation: test retrieval before judging the final answer
Create a document-grounded evaluation dataset that separates retrieval quality, answer support and safe refusal.
Read moreHow to design a reproducible local LLM test
A measurement protocol for comparing local inference without inventing benchmark results.
Read moreHow to secure AI agents in production: put policy outside the prompt
A practical agent security design with tool permissions, exact-action approvals, untrusted data handling and failure-path tests.
Read moreA useful single-page SEO audit: evidence, fixes and limits
Understand what a public HTML audit can establish and how to turn its findings into a prioritized implementation plan.
Read moreStructured JSON from local models: validate more than syntax
Build a model-output contract with schema validation, semantic checks and safe failure handling before integrating writes.
Read moreNo matching content. Try a broader search.