https://aayushgoel.dev/blog/ Notes on backend systems that have to survive production — ERP integration, messaging infrastructure, and the unglamorous work of making them reliable. en mail@aayushgoel.dev (Aayush Goel) mail@aayushgoel.dev (Aayush Goel) Thu, 05 Nov 2026 09:00:00 +0530 https://aayushgoel.dev/blog/latency-in-100ms-steps https://aayushgoel.dev/blog/latency-in-100ms-steps Thu, 05 Nov 2026 09:00:00 +0530 mail@aayushgoel.dev (Aayush Goel) Engineering I spent a day tuning Elasticsearch queries. The slowdown was CPU throttling from our VPS host, triggered by a Logstash container that had crash-looped 41,687 times. https://aayushgoel.dev/blog/too-slow-is-too-big https://aayushgoel.dev/blog/too-slow-is-too-big Mon, 26 Oct 2026 09:00:00 +0530 mail@aayushgoel.dev (Aayush Goel) Engineering My fallback for large Tally exports waited for a "too big" error. The biggest company never sent one. It timed out at Cloudflare's 120 seconds instead, and the fallback never ran. https://aayushgoel.dev/blog/the-floor-was-the-call https://aayushgoel.dev/blog/the-floor-was-the-call Sat, 24 Oct 2026 09:00:00 +0530 mail@aayushgoel.dev (Aayush Goel) Engineering An LLM classifier took three seconds to answer even "confirm". Shrinking the output barely helped, because the model's fixed per-call cost was the floor. A faster model was right there, and I still didn't take it. https://aayushgoel.dev/blog/five-agents-three-approval-gates https://aayushgoel.dev/blog/five-agents-three-approval-gates Tue, 22 Sep 2026 09:00:00 +0530 mail@aayushgoel.dev (Aayush Goel) Engineering A multi-agent LLM system that spends real money on ads: every ad created paused, three human approval gates, and every model call logged. https://aayushgoel.dev/blog/ten-seconds-to-two https://aayushgoel.dev/blog/ten-seconds-to-two Tue, 22 Sep 2026 09:00:00 +0530 mail@aayushgoel.dev (Aayush Goel) Engineering A WhatsApp chatbot took ten seconds to reply. Twelve serial reads, a blocking typing indicator and an LLM call on the critical path. Getting it to two. https://aayushgoel.dev/blog/the-index-postgres-quietly-stopped-using https://aayushgoel.dev/blog/the-index-postgres-quietly-stopped-using Tue, 22 Sep 2026 09:00:00 +0530 mail@aayushgoel.dev (Aayush Goel) Engineering A refactor made catalogue search 54% slower at p90. The partial index no longer matched the query, so Postgres quietly skipped it. 281 ms to 29 ms. https://aayushgoel.dev/blog/your-timeout-is-not-a-timeout https://aayushgoel.dev/blog/your-timeout-is-not-a-timeout Tue, 22 Sep 2026 09:00:00 +0530 mail@aayushgoel.dev (Aayush Goel) Engineering axios' timeout is a socket-idle timeout, not a deadline. A trickling response never goes idle, so it never fires. Plus a 26 MB parse that outran its own timer. https://aayushgoel.dev/blog/multi-tenant-whatsapp-cloud-api https://aayushgoel.dev/blog/multi-tenant-whatsapp-cloud-api Mon, 21 Sep 2026 09:00:00 +0530 mail@aayushgoel.dev (Aayush Goel) Engineering One WhatsApp Cloud API number serving many stores: resolving the tenant from a webhook that doesn't name one, and caching it without guessing wrong. https://aayushgoel.dev/blog/tally-erp-two-way-sync https://aayushgoel.dev/blog/tally-erp-two-way-sync Mon, 21 Sep 2026 09:00:00 +0530 mail@aayushgoel.dev (Aayush Goel) Engineering Tally has no REST API, answers HTTP 200 on failure, and its documented batching is a no-op. What two generations of sync taught me.