Field notes on production GenAI and backend engineering, plus personal essays.https://anupam.info/en-inhttps://anupam.info/articles/building-a-grounded-portfolio-chatbot.htmlhttps://anupam.info/articles/building-a-grounded-portfolio-chatbot.htmlHow the "Ask Anupam's AI" widget works: a build-time chunk index over the whole site, dependency-free BM25 retrieval (no embeddings API), page-aware boosting, Groq-then-Gemini streaming over SSE, the guardrails that keep it from inventing facts, and three real incidents from the week it went live.Sun, 27 Sep 2026 00:00:00 GMTRAGLLMChatbotsProduction Engineeringhttps://anupam.info/articles/my-portfolio-has-an-mcp-server.htmlhttps://anupam.info/articles/my-portfolio-has-an-mcp-server.htmlA public, read-only Model Context Protocol server for this site: the 8 tools it exposes, why it targets protocol version 2025-06-18 instead of the newest one, why it's hand-rolled JSON-RPC instead of the official SDK on a serverless route, and how to connect Claude, Claude Code, Cursor, or curl to it.Sun, 27 Sep 2026 00:00:00 GMTMCPAI AgentsProtocol Designhttps://anupam.info/til/checksum-dedup-uploadshttps://anupam.info/til/checksum-dedup-uploadsTwo uploads of the same PDF shouldn't cost two embedding runs - hash the file up front and short-circuit indexing if that hash is already in the system.Wed, 23 Sep 2026 00:00:00 GMTTILRAGPostgreSQLBackendhttps://anupam.info/til/graceful-sigterm-queue-consumershttps://anupam.info/til/graceful-sigterm-queue-consumersKubernetes sends SIGTERM before killing a pod, and a queue consumer that ignores it drops whatever message it was mid-processing - catch it, finish the current message, then exit.Wed, 23 Sep 2026 00:00:00 GMTTILKubernetesReliabilityQueueshttps://anupam.info/til/file-hash-change-detection-cihttps://anupam.info/til/file-hash-change-detection-ciRedeploying every Lambda function on every push doesn't scale past a couple dozen functions - hash each function's source and only redeploy the ones whose hash actually changed.Wed, 23 Sep 2026 00:00:00 GMTTILAWS LambdaCI/CDCloudFormationhttps://anupam.info/til/model-independent-agentshttps://anupam.info/til/model-independent-agentsHardcoding a model's quirks into agent logic means every model upgrade risks silently changing behaviour - push model-specific details to config and keep the orchestration model-agnostic.Wed, 23 Sep 2026 00:00:00 GMTTILAgentsLLMPrompt Engineeringhttps://anupam.info/til/presigned-urls-past-payload-limithttps://anupam.info/til/presigned-urls-past-payload-limitAmazon API Gateway enforces a 10 MB payload limit on REST APIs, and it isn't configurable - route the file straight to S3 with a presigned URL instead of proxying it through Lambda.Wed, 23 Sep 2026 00:00:00 GMTTILAWSAPI GatewayS3Lambdahttps://anupam.info/til/token-aware-embedding-batchinghttps://anupam.info/til/token-aware-embedding-batchingBatching embedding requests by a fixed item count still hits 429s once individual chunks vary in length - batch by token budget instead, and back off on Retry-After.Wed, 23 Sep 2026 00:00:00 GMTTILAzure OpenAIEmbeddingsReliabilityRAGhttps://anupam.info/til/stale-job-reaperhttps://anupam.info/til/stale-job-reaperA job marked 'processing' by a worker that then crashed or got OOM-killed stays 'processing' forever unless something else notices and resets it.Wed, 23 Sep 2026 00:00:00 GMTTILPipelinesReliabilityPostgreSQLhttps://anupam.info/til/docx-to-pdf-libreoffice-fallbackhttps://anupam.info/til/docx-to-pdf-libreoffice-fallbackA document-intelligence service that reads PDFs cleanly can still choke on certain DOCX files - convert to PDF first with headless LibreOffice, then retry, instead of failing the upload outright.Wed, 23 Sep 2026 00:00:00 GMTTILDocument AIAzureBackendhttps://anupam.info/articles/taming-the-dice.htmlhttps://anupam.info/articles/taming-the-dice.htmlThe complete field guide to LLM reliability, written for everyone from zero background to production. How language models actually work, why the same prompt gives different answers even at temperature 0 (including the 2025 batch-invariance finding), temperature vs top_p vs top_k defined with worked numbers, what seed and system_fingerprint really buy you, model-agnostic prompt engineering, structured outputs and constrained decoding, how to test any of it, and a tour of every other way these systems bite in production. Seven diagrams, thirty-one cited sources.Tue, 25 Aug 2026 00:00:00 GMTLLMDeterminismPrompt EngineeringReliabilityFor Everyonehttps://anupam.info/articles/why-your-rag-pipeline-is-confidently-wrong.htmlhttps://anupam.info/articles/why-your-rag-pipeline-is-confidently-wrong.htmlThe failure modes that never show up in a demo - silent retrieval misses, semantic near-misses, stale embeddings, citation-answer mismatch - why "it worked in testing" is a trap without a retrieval trace, and a concrete reliability playbook from two production RAG systems.Thu, 23 Jul 2026 00:00:00 GMTRAGLLMProduction Engineeringhttps://anupam.info/perspectives/the-illusion-of-safety.htmlhttps://anupam.info/perspectives/the-illusion-of-safety.htmlOn the wall we build around our own comfort, the thin membrane that actually protects it, and what's still ours to control once that wall finally tears - plus a concrete idea for accountability that's hard to silence.Sat, 18 Jul 2026 00:00:00 GMTPerspectivehttps://anupam.info/articles/llm-determinism-and-trust.htmlhttps://anupam.info/articles/llm-determinism-and-trust.htmlWhy LLM output determinism is a legal problem, what Azure OpenAI's seed and system_fingerprint actually guarantee (less than you'd hope), why AWS Bedrock has no equivalent for text models, a practical audit-logging playbook, and an honest opinion on building on something that's fundamentally guessing the next word. Scrollable field notes with real vendor docs cited.Sun, 12 Jul 2026 00:00:00 GMTLLMReliabilityLegal Techhttps://anupam.info/articles/war-of-the-agents.htmlhttps://anupam.info/articles/war-of-the-agents.htmlA deep dive into GPT-5.6 Sol Ultra, Claude 5 Fable/Mythos, and Gemini 3.1. We look past the heavily gamed Terminal-Bench 2.1 scores to see what these frontier models actually mean for everyday developers, and why the community is pushing back on "safety-gated" reasoning capabilities.Mon, 29 Jun 2026 00:00:00 GMTAI AgentsGPT-5.6Claude 5Gemini 3.1Industry Opinionhttps://anupam.info/articles/everyday-ai-playbook.htmlhttps://anupam.info/articles/everyday-ai-playbook.htmlHow to make AI save you real hours - whether you're a student, a doctor, a researcher, running a business, or just getting through your to-do list. A developer's field guide to the durable habits behind good AI use: the right mindset, per-track use cases, the two skills that decide whether any of it works, and how to stay out of the trap where AI quietly costs you more than it saves. Scrollable, illustrated guide.Fri, 19 Jun 2026 00:00:00 GMTAIProductivityFor Everyonehttps://anupam.info/articles/maximum-output-with-claude-code.htmlhttps://anupam.info/articles/maximum-output-with-claude-code.htmlA field guide to running AI like a system, not a chat box - persistent context, a research-first planning gate, controlled execution with subagents, and the two human checkpoints that catch what automation can't. Scrollable, illustrated walkthrough.Thu, 18 Jun 2026 00:00:00 GMTClaude CodeAI WorkflowProductivity