THE ENGINEERING JOURNAL
Ideas, tested
against the work.
Notes on building useful AI systems, evaluating tradeoffs and thinking across the stack.
Explore the technical publication.
Architecture decisions, experiments and lessons from building AI systems.
LATEST / ENGINEERING
17 NOVEMBER 2026 / 7 MIN READCoverage complete.
Evidence still needs review.
29 untouched classifications tested. Recovered failures, preserved original results and modest AI-rated evidence differences with criteria caveats.
Read the completeness continuation 17 NOVEMBER 2026 / 7 MIN READEvidence completeness.
Evaluation interrupted.
Five completed pairs, no observed gain. One HTTP 429, 29 unattempted classifications and the limits of blinded AI evidence review.
Read the evidence-completeness evaluation 17 NOVEMBER 2026 / 9 MIN READSensitive context.
Original failures preserved.
A stopped evaluation and twelve-case continuation. The combined result, medical review miss, original failures and AI-review limits.
Read the sensitive-context evaluation 17 NOVEMBER 2026 / 9 MIN READRecovered response.
Unchanged correctness.
One HTTP 503 retry restored a valid but wrong route. All 40 cases, both review misses and the limits of AI review.
Read the reliability evaluation 17 NOVEMBER 2026 / 9 MIN READNegated closure.
No observed gain.
Identical routing on 35 common-valid cases, five failed calls and an unmeasured correction. The complete comparison and AI-review limits.
Read the negated-closure evaluation 16 NOVEMBER 2026 / 9 MIN READVague requests.
Clearer review.
Three observed routing corrections, two failed calls and a shared unnecessary review. The findings and limits of a fresh AI-reviewed comparison.
Read the vague-request evaluation 16 NOVEMBER 2026 / 9 MIN READBounded evidence.
No observed gain.
Zero contract rejections in both modes, seven failed calls and the same review miss. What exact extraction does—and does not—establish.
Read the evidence evaluation 16 NOVEMBER 2026 / 9 MIN READClearer routing.
Visible failures.
Five observed routing corrections, eight failed calls and what separate AI label review can—and cannot—establish.
Read the paired evaluation 16 NOVEMBER 2026 / 7 MIN READValid output.
Missed review.
All 40 outputs passed validation. Three mapped review cases were missed. Diagnostics, independent inputs and the limits of the labels.
Read the findings 16 NOVEMBER 2026 / 7 MIN READGemini improved triage.
The failures still matter.
38 of 40 exact decisions, one rejected output and the routing coverage lost when rules constrain the model.
Read the comparison 16 NOVEMBER 2026 / 7 MIN READWhat 30 triage
cases revealed.
25 exact matches, five documented failures and the review misses hidden by an accuracy headline.
Read the evaluation 16 NOVEMBER 2026 / 6 MIN READFrom AI demo to production:
start with the workflow.
A practical framework for defining scope, testing quality, integrating tools and owning the system after launch.
Read the articleInspect the triage workflow.
Try synthetic requests and see route proposals, review decisions and their reasons. Rules baseline; no live model or business actions.
How we would evaluate inference efficiency.
Quality, latency, throughput, memory and cost, recorded together.
START WITH A REAL PROBLEM
What should AI
make possible for you?
Bring the workflow. We’ll help define the next step.
Request a workflow assessmentWORKFLOW DESIGN NOTE
What happens after an inquiry?
A synthetic lead-to-revenue walkthrough, with approval paths, failure handling requirements and a scorecard for the first pilot.
Explore the walkthrough and evaluation plan ↗