arpit agrawalPersonal website of Arpit Agrawal, senior ML platform engineer. Post-mortems, decision records and notes from running ML inference infrastructure.https://arpitagrawal.devThe night Spot capacity disappearedhttps://arpitagrawal.dev/docs/pm4https://arpitagrawal.dev/docs/pm4Two regions, forty quiet minutes, and a fallback path that never triggered.Fri, 14 Aug 2026 00:00:00 GMTpost-mortemTwo-phase commit for credits, not sagashttps://arpitagrawal.dev/docs/adr11https://arpitagrawal.dev/docs/adr11Why the billing system holds a reservation before any GPU work starts.Fri, 12 Jun 2026 00:00:00 GMTdecisionWhat 100 TensorRT conversions taught me about ONNXhttps://arpitagrawal.dev/docs/n19https://arpitagrawal.dev/docs/n19Most failures happened before TensorRT was involved.Thu, 09 Apr 2026 00:00:00 GMTnoteRouting GPU work by complexity instead of queue depthhttps://arpitagrawal.dev/docs/adr9https://arpitagrawal.dev/docs/adr9Queue depth measures how much work is waiting. It says nothing about how heavy that work is.Fri, 06 Feb 2026 00:00:00 GMTdecisionNotes from mentoring two engineershttps://arpitagrawal.dev/docs/n14https://arpitagrawal.dev/docs/n14What changed when I stopped answering questions straight away.Thu, 04 Dec 2025 00:00:00 GMTnoteA NATS consumer that acknowledged too earlyhttps://arpitagrawal.dev/docs/pm3https://arpitagrawal.dev/docs/pm3The acknowledgement went out before the work had actually finished.Sun, 02 Nov 2025 00:00:00 GMTpost-mortemWhy the platform runs on Karpenterhttps://arpitagrawal.dev/docs/adr7https://arpitagrawal.dev/docs/adr7Choosing a node autoscaler for bursty, GPU-heavy workloads.Mon, 18 Aug 2025 00:00:00 GMTdecisionWhat I kept from CMU's scalable systems trackhttps://arpitagrawal.dev/docs/n9https://arpitagrawal.dev/docs/n9Two ideas I still use every week.Wed, 21 May 2025 00:00:00 GMTnote