Writing for operators + engineers
Practical notes on
building less manual work.
Architecture, reliability, process mapping, and economics. No generic “AI trends” and no beginner tool roundups.
All articles
Field notes you can inspect.
Written to help a buyer make a better process decision and an engineer inspect the implementation logic.
Getting Close to Frontier Results With Open-Source Models, One Task at a Time
Open-source LLMs vs frontier models on one task: a reproducible method for finding out how close a system of open models gets, and what it costs to close the rest of the gap.
Read articleQ4_K_M vs Q8_0: What Quantization Costs You
Q4_K_M vs Q8_0 and the other GGUF quantization labels: how to estimate memory and speed from them, and how much quality each level loses, using llama.cpp's own measurements.
Read articleMixture of Experts: Total Parameters for Memory, Active Parameters for Speed
Why an MoE model needs the memory of its total size but runs at the speed of its active size, and how to run one on a small GPU.
Read articleSmall Task-Specific Models That Beat Big Ones at One Job
Four task-specific open models for entity extraction, structured extraction, reranking and speech, and when a specialist is the better choice.
Read articleDo You Need a Vector Database? What to Get Right First
pgvector or a dedicated vector database: exact search, ANN indexes, the filtering problem and metadata decide retrieval quality, and Postgres is often enough.
Read articlePostgres as a Queue: SKIP LOCKED and the Transactional Outbox
A jobs table with SKIP LOCKED and a transactional outbox close the gap between saving work and scheduling it, which is where AI workflows lose items.
Read articleDisable Thinking Mode When You Want Ideas
Why reasoning mode can narrow a brainstorm, how to disable thinking in Qwen3 and Ollama, and an experiment to measure the effect on your own prompts.
Read articleWhat LLM Temperature Actually Does
LLM temperature is one division before softmax: how it reshapes the next-token distribution, how it interacts with top-p and min-p, and why temperature 0 isn't deterministic.
Read articleMost AI Automations Should Not Be Agents
A reliable operational workflow is usually mostly ordinary software, with a small number of bounded AI steps where language or ambiguity makes rules brittle.
Read articleWhat Production n8n Workflows Need That Tutorials Usually Skip
The canvas is only the visible part. Production readiness lives in execution identity, failure semantics, ownership and the ability to reconcile side effects.
Read articleDesigning AI Workflows for Companies That Cannot Afford Silent Errors
Treat model output as untrusted input, place controls around it, and make uncertainty an explicit workflow state.
Read articleReliable LLM Extraction From Messy Business Documents
A production extraction system is a pipeline of document controls, schema design, validation, evaluation and exception handling—not one prompt.
Read articleHow I Decide Whether an Automation Is Worth Building
A transparent capacity and risk model beats a list of features. Here is the framework I use before recommending implementation.
Read articlen8n vs Custom Code: The Decision Framework I Use
The useful answer is rarely ideological. Choose the representation that makes the workflow easiest to change, operate and trust.
Read articleHuman-in-the-Loop Is a Product Feature, Not an AI Failure
A good approval step concentrates human judgment at the consequence boundary and removes everything the reviewer should not have to do.
Read articleThe Hidden Engineering Problem in AI Automation: Idempotency
Webhook retries, worker crashes and lost responses can repeat a business action. Preventing duplicates requires a business key and state—not a unique workflow run.
Read articleHow I Evaluate AI Steps Inside Deterministic Workflows
You can test a classifier or extractor rigorously without turning the whole process into an agent benchmark.
Read articleFrom SOP to Software: How I Map a Manual Process Before Automating It
The written procedure is a starting hypothesis. The real workflow lives in triggers, handoffs, exceptions and the shortcuts experienced operators use.
Read articlePrefer a concrete system?
Every principle is applied in the design studies.
See before/after process maps, action boundaries, error paths, and the validation plan each study proposes.
Built against a documented reference workload when published; not presented as client work. Numbers appear only with method, date, and artifact.