https://shinyq.my.id/en/feed.xmlArticles by Kurniadi Ahmad Wijaya2026-11-02T00:00:00ZKurniadi Ahmad Wijayahttps://shinyq.my.id/enhttps://shinyq.my.id/en/blog/ten-very-hard-benchmark-tasks-in-ten-daysTen Very Hard Benchmark Tasks in Ten Days2026-11-02T00:00:00Z2026-11-02T00:00:00ZFirst AI benchmark job: ten Indonesian tasks in ten days, half merged, most rated very hard, the top band. Over 300 ideas killed. The method, not the tasks.https://shinyq.my.id/en/blog/aria-acs-to-livekitMoving ARIA From ACS to LiveKit2026-10-27T00:00:00Z2026-10-27T00:00:00ZARIA moved from ACS to self-hosted LiveKit. A PoC drill measured 425 ms median hand-back over 20 rounds; no matched ACS-to-LiveKit speedup was measured.https://shinyq.my.id/en/blog/untrusted-code-next-to-production-secretsUntrusted Code Next to Production Secrets2026-10-24T00:00:00Z2026-10-24T00:00:00ZAn audit showed a repository's code could run inside the API holding production secrets. Same-day stopgap, then isolated workers and a 9-minute outage.https://shinyq.my.id/en/blog/the-stall-that-was-never-in-the-queryThe Stall That Was Never in the Query2026-10-24T00:00:00Z2026-10-24T00:00:00ZA Fabric dashboard took ~21 s per request, static files included. A zero-row probe showed capacity throttling, not query cost. I had blamed indexes first.https://shinyq.my.id/en/blog/the-lock-that-protected-the-install-also-blocked-the-appThe Install Lock That Also Blocked the App2026-10-24T00:00:00Z2026-10-24T00:00:00ZIn the first managed application test installs, every AI call failed with correct roles. The lock blocked the app's own identities. 36 defects followed.https://shinyq.my.id/en/blog/clearbox-inbox-cleanup-with-microsoft-decision-1Clearbox: Triage with Microsoft-Decision-12026-10-16T00:00:00Z2026-10-16T00:00:00ZClearbox pre-sorts Gmail and Outlook email for approval, using Microsoft-Decision-1 and guardrails. The evaluation covered 40 synthetic emails, not an inbox.https://shinyq.my.id/en/blog/making-a-voice-agent-interruptibleMaking a Voice Agent Interruptible2026-10-09T00:00:00Z2026-10-09T00:00:00ZVoice-agent barge-in reached ~0.2 s in a controlled knowledge-search harness, n=4 calls per arm. Later live cut-off tests showed no improvement.https://shinyq.my.id/en/blog/why-the-better-fraud-engine-shipped-in-shadow-modeWhy the Fraud Engine Stayed in Shadow Mode2026-10-09T00:00:00Z2026-10-09T00:00:00ZIn offline tests on the noisy-clean development slice, FPR fell 36.4% to 9.1%, but High-risk recall fell. The engine shipped only in shadow mode.https://shinyq.my.id/en/blog/45-percent-of-our-llm-calls-failedWhy 45% of a CV Screening Batch Failed2026-10-09T00:00:00Z2026-10-09T00:00:00ZAbout 45% of 22 CV/job pairs failed. Two parsing bugs were fixed with seven regression tests, but no rerun of the original batch was recorded.https://shinyq.my.id/en/blog/spec-driven-development-with-ai-coding-agentsSpec-Driven Development with AI Coding Agents2026-10-09T00:00:00Z2026-10-09T00:00:00ZAI agents wrote most of a solo-built vendor portal. CI missed a skipped integration tier and two sign-in bugs. Productivity gains were not measured.