Engineering notes
Practical engineering notes on embedded validation, agentic systems, and machine learning.
Twenty in-depth articles on building reliable automation, reviewable AI workflows, and maintainable ML systems.
Featured · Jul 29, 2026
Robotic Validation Is a Closed-Loop Evidence System
A useful robotic test rig does more than repeat physical actions. It controls the product, observes the response, decides against explicit criteria, and preserves enough evidence to explain every result.
Jul 28, 2026
Choosing the Right Embedded Validation Layer: SIL, HIL, Bench, Vision, or Robot?
The best embedded validation strategy is not the most realistic test environment. It is a deliberate portfolio that finds each class of problem at the cheapest layer capable of exposing it.
Jul 27, 2026
Designing a Device-Control Layer That Recovers Instead of Failing Randomly
Reliable test automation begins below the test cases. A state-aware device-control layer turns fragile scripts into bounded operations with explicit readiness, timeouts, recovery, and evidence.
Jul 26, 2026
Evidence-First Test Automation: What Every Run Should Leave Behind
A green dashboard is not enough. Useful automation preserves the context, observations, timing, and artifacts needed to trust a pass and diagnose a failure.
Jul 25, 2026
Agentic Workflows Need Boundaries, Not More Autonomy
The useful question is not how much an AI agent can do, but which actions it may take, under what conditions, with what evidence, and who remains accountable.
Jul 24, 2026
A Practical RAG Pipeline Starts Before Retrieval
Retrieval-augmented generation is not an embeddings feature. It is a controlled path from approved sources to a grounded answer, with access, freshness, and evaluation designed into every stage.
Jul 23, 2026
How to Evaluate an Agentic Workflow Before It Touches Production
An agent is not ready because its answers look plausible. Evaluate the complete workflow across task quality, grounding, tool use, permissions, recovery, and human review effort.
Jul 22, 2026
Before Training a Custom ML Model, Prove the Problem Is Learnable
A feasibility study should test the decision, labels, data, baseline, error costs, and deployment constraints before a team invests in a custom model.
Jul 21, 2026
Edge AI Is a System Trade-Off, Not a Model Export
A model that runs on a workstation is not automatically an edge solution. Hardware, latency, memory, power, thermals, preprocessing, updates, and recovery must be designed together.
Jul 20, 2026
Computer Vision Fails in Conditions, Not in Averages
An overall score can hide the exact lighting, motion, device, or product condition that breaks a vision system. Evaluate the operating envelope as a matrix of real conditions.
Jul 19, 2026
When the Test Rig Lies: Separating Product Failures from Infrastructure Failures
A failed test is only useful when the automation can distinguish a product defect from a broken cable, stale camera, exhausted relay, noisy supply, or unhealthy host.
Jul 18, 2026
Power Cycling an Embedded Device Is a State Machine, Not a Relay Toggle
Reliable power-cycle testing controls discharge, rail timing, reset, boot observation, communication readiness, and recovery instead of treating OFF and ON as two relay commands.
Jul 17, 2026
Testing Timing Behaviour Without Creating Flaky Tests
Timing tests become trustworthy when they define the event pair, clock, operating conditions, distribution, and decision rule instead of relying on one wall-clock measurement.
Jul 16, 2026
Design Embedded Products for Testability Before the First Prototype
A few deliberate choices in power, interfaces, diagnostics, trace, fixtures, and software state can turn validation from reverse engineering into a maintainable product capability.
Jul 15, 2026
A Robot Framework Architecture That Scales Beyond One Test Suite
Scalable Robot Framework projects keep business-readable tests, reusable domain keywords, state-aware Python libraries, configuration, and evidence responsibilities separate.
Jul 14, 2026
Flaky Tests Are a Classification Problem, Not a Retry Setting
Retries can measure reproducibility, but they cannot tell whether instability comes from the test, framework, product, dependency, operating system, network, or hardware.
Jul 13, 2026
Prompt Injection Is a System-Boundary Failure
Treat retrieved documents, messages, webpages, and model output as untrusted data. Security comes from separating instructions, evidence, authority, and deterministic enforcement.
Jul 12, 2026
Human Approval That Actually Controls an Agentic Workflow
An approval gate is meaningful only when the reviewer sees the exact action, evidence, uncertainty, and consequence—and the system executes nothing materially different afterward.
Jul 11, 2026
Data Drift, Concept Drift, and Sensor Drift Are Different Problems
A changed input distribution, a changed relationship between input and outcome, and a degraded measurement system may look similar in a dashboard but require different evidence and remedies.
Jul 10, 2026
Monitoring an ML System When Ground Truth Arrives Late—or Never
When outcome labels are delayed, monitor the input contract, data pipeline, prediction behaviour, uncertainty, system health, review signals, and business process without pretending those proxies equal accuracy.