Monitoring: Software Signals, Checks, and Useful Alerts
A checked HTTP example shows why process health can stay green during a failed checkout, with practical signals and an alert checklist.
AI, software engineering, and startup insights
Practical guides and evidence-led analysis for developers and founders, covering AI risks, software systems, and building technology businesses.
Latest Story
Race Condition: Reproduce a Lost Update and Fix the Lock Scope
Force a lost update under CPython 3.12.14, compare shared and failed lock scopes, and identify what the small test does not prove.
Latest news
Multi Version Concurrency Control: PostgreSQL Snapshots and Locks
Predict which row a query sees, reproduce two-session PostgreSQL tests, and distinguish snapshots, writer conflicts, retries, and cleanup.
Sales Conversion Rate: Fix the Denominator Before Comparing
Audit a small lead cohort to see why two valid rates differ, which sales events to exclude, and how to keep comparisons fair.
Model Drift: Detect Failing AI Predictions and Choose a Response
Use the Google Flu Trends failure to distinguish changing inputs from measured prediction loss, then record the evidence needed for a response.
Prioritization: A Product Backlog Decision Worksheet
Use a checked five-item backlog to separate RICE rankings from dependencies, capacity and missing evidence, then record the choice and review trigger.
Serialization: Formats, Round Trips, and Data Contracts
A valid JSON payload can still change an ID or a type. See checked Python and JavaScript cases, then define what your consumer must preserve.
Proof of Concept Testing: A Startup Test Plan and Decision Gate
Build a finite feasibility test with clear gates, a cost cap and a worked example where a sample pass still leads to a failed decision.
Product Differentiation: A SaaS Claim and Evidence Worksheet
Use a fictional SaaS worksheet to separate a product capability, buyer relevance, and a fair comparison before choosing claim wording.
AI Data Poisoning: Trace Training Data and Its Limits
Trace AI data poisoning through training data and source changes. Use a checked custody example, clear evidence limits, and practical response decisions.