Two AI Jobs, One Messy Log: Keep Every Run in Its Own Lane
Reproduce a concurrent run-ID mix-up, then keep task-local event labels with Python ContextVar. Includes ten tests, cancellation cleanup and a child-task inheritance check.
BY JASON EAD
AI engineering & business applications
Original guides and analysis about coding tools, models, GitHub projects, agents, and creative AI workflows.
Published articles
37 articles
Drowning in AI Images? Build a Contact Sheet You Can Actually Review
Build an uncropped, labeled contact sheet and source manifest for campaign candidates. A tested local Pillow workflow keeps creative reviews tied to the right files.
Why Your AI Quote Has Two Different Totals
Make decimal input and rounding stages explicit in an AI quote handoff. Reproduce the one-cent policy difference and 13 runtime tests before adapting the calculation.
Git Can Hide Part of Your AI Coding Handoff
Find intended files that ordinary Git status can omit. Reproduce indexed, ignored, untracked and missing handoff states with seven isolated checks.
Stop a Data Join From Cloning Your AI Briefs
Catch duplicate lookup keys and unmatched accounts before a data join multiplies your AI workflow records. Reproduce the pandas example with 12 runtime tests.
Stop Guessing Your AI Promo’s Audio Level
Measure, normalize and re-measure a saved WAV with FFmpeg. Reproduce a two-pass audio drill with five real checks before adapting it to your promo workflow.
Your AI Scored 100%. Half the Evaluation Is Missing.
Audit missing model-evaluation grades before trusting a perfect score. Reproduce a CSV coverage report with 11 tests and explicit denominators.
Your AI Refactor Passed Tests. This Name Still Does Not Exist.
Reproduce a passing thin test suite, an F821 finding and a real NameError, then repair the code and test the missing behavior before accepting a coding task.
Your AI Needs Sources. Start With Search You Can Inspect.
Build a local SQLite FTS5 procedure-search baseline. Reproduce useful keyword hits and a revealing synonym miss before comparing more complex retrieval.
Your AI JSON Has Two Answers. Python Quietly Picked One.
Reject duplicate JSON keys before a default parser silently keeps the last value. Test nested and escaped duplicates, non-finite values and the raw-text boundary.
Your AI Video Exported. Is the File Actually Deliverable?
Check the exported video, not just the progress bar. Test dimensions, pixel shape, duration and required audio, then decode selected streams before handoff.
Your AI Says 1:30 A.M. Which One?
Catch repeated and nonexistent local times before an AI-assisted request reaches a calendar. Build a tested New York conversion boundary with explicit human choices.
Your AI App Runs. Its Dependencies Might Still Be Wrong.
Reproduce a stale uv lockfile, see why frozen can run the old dependency, and give your coding assistant a consistency check it can prove.
Your Cheapest AI Model Might Be Your Most Expensive Workflow
Include human review time when comparing AI workflows. Use a tested decimal calculator and a task-trial record, with clearly fictional inputs and no model-performance claims.
Your AI CSV Could Be a Spreadsheet Trap
Inspect parsed CSV cells for suspicious prefixes before spreadsheet delivery. Includes a tested text-only review gate, false-positive examples and clear limits.
Your AI App Broke. Find the Commit Before Rewriting It.
Use a repeatable business test and Git bisect to locate an injected quote-calculation regression in a disposable four-commit repository.
Your Agent Skill Is Not Permission to Hit Send
Package a draft-only business reply skill with clear stopping points, source-backed review packets and five boundary tests before enabling delivery.
Stop Blank AI Reports Before They Reach Your Clients
Build a tested Jinja rendering boundary that rejects missing report fields, escapes HTML text and keeps human approval separate from delivery.