https://sohanjagadeesh.com/writing Software, stories, systems, and working notes from Sohan Jagadeesh. en-au Mon, 20 Jul 2026 00:00:00 GMT https://sohanjagadeesh.com/writing/ship-model-changes-behind-evaluations-and-rollback https://sohanjagadeesh.com/writing/ship-model-changes-behind-evaluations-and-rollback Mon, 20 Jul 2026 00:00:00 GMT A release discipline for prompts, models, retrieval, tools, and policies that can regress without a code compile failing. AI Delivery Evaluation https://sohanjagadeesh.com/writing/streaming-is-a-product-state-machine https://sohanjagadeesh.com/writing/streaming-is-a-product-state-machine Mon, 20 Jul 2026 00:00:00 GMT A UI and systems guide to partial output, tool pauses, validation, cancellation, accessibility, and truthful terminal states. AI UX Architecture https://sohanjagadeesh.com/writing/separate-instructions-evidence-and-user-input https://sohanjagadeesh.com/writing/separate-instructions-evidence-and-user-input Mon, 20 Jul 2026 00:00:00 GMT A trust-ordered prompt architecture that keeps product policy, user intent, retrieved material, and tool results visibly distinct. AI Security Architecture https://sohanjagadeesh.com/writing/red-team-the-seams-not-only-the-chatbot https://sohanjagadeesh.com/writing/red-team-the-seams-not-only-the-chatbot Mon, 20 Jul 2026 00:00:00 GMT A scenario-led red-team plan for retrieval, tools, memory, permissions, streaming, and recovery around the model. AI Security Evaluation https://sohanjagadeesh.com/writing/freshness-is-a-retrieval-requirement https://sohanjagadeesh.com/writing/freshness-is-a-retrieval-requirement Mon, 20 Jul 2026 00:00:00 GMT A version-aware retrieval design for detecting change, modeling validity, removing stale evidence, and citing the right point in time. AI Retrieval Data https://sohanjagadeesh.com/writing/give-genai-systems-service-level-objectives https://sohanjagadeesh.com/writing/give-genai-systems-service-level-objectives Mon, 20 Jul 2026 00:00:00 GMT How to define reliability targets for useful, grounded outcomes instead of measuring only API uptime. AI Reliability Observability https://sohanjagadeesh.com/writing/build-a-privacy-boundary-before-the-prompt https://sohanjagadeesh.com/writing/build-a-privacy-boundary-before-the-prompt Mon, 20 Jul 2026 00:00:00 GMT A data-flow method for minimizing, classifying, redacting, and governing information before it reaches a model or trace. AI Privacy Architecture https://sohanjagadeesh.com/writing/design-multimodal-ingestion-for-evidence https://sohanjagadeesh.com/writing/design-multimodal-ingestion-for-evidence Mon, 20 Jul 2026 00:00:00 GMT A document pipeline that preserves pages, layout, tables, images, and provenance instead of flattening everything into unreliable text. AI Retrieval Documents https://sohanjagadeesh.com/writing/agent-memory-is-product-state https://sohanjagadeesh.com/writing/agent-memory-is-product-state Mon, 20 Jul 2026 00:00:00 GMT A design model for consent, provenance, correction, expiry, and retrieval when an agent is expected to remember. AI Agents Architecture https://sohanjagadeesh.com/writing/put-human-approval-at-irreversible-boundaries https://sohanjagadeesh.com/writing/put-human-approval-at-irreversible-boundaries Mon, 20 Jul 2026 00:00:00 GMT How to place approvals where they preserve control, show the real diff, and survive long-running agent workflows. AI Agents Safety https://sohanjagadeesh.com/writing/route-models-by-task-not-prestige https://sohanjagadeesh.com/writing/route-models-by-task-not-prestige Mon, 20 Jul 2026 00:00:00 GMT A measured routing strategy for balancing task success, cost, latency, and fallback without building a hidden leaderboard. AI Models Architecture https://sohanjagadeesh.com/writing/budget-latency-before-optimizing-tokens https://sohanjagadeesh.com/writing/budget-latency-before-optimizing-tokens Mon, 20 Jul 2026 00:00:00 GMT A systems approach to time-to-first-useful-result, tail latency, caching, and the pauses users actually feel. AI Performance Architecture https://sohanjagadeesh.com/writing/trace-the-whole-inference-path https://sohanjagadeesh.com/writing/trace-the-whole-inference-path Mon, 20 Jul 2026 00:00:00 GMT An observability map for following a request across retrieval, model calls, tools, validation, and the final user experience. AI Observability Architecture https://sohanjagadeesh.com/writing/calibrate-llm-as-a-judge https://sohanjagadeesh.com/writing/calibrate-llm-as-a-judge Mon, 20 Jul 2026 00:00:00 GMT A practical protocol for using model graders without confusing scalable judgement with objective truth. AI Evaluation Quality https://sohanjagadeesh.com/writing/tool-calls-need-contracts https://sohanjagadeesh.com/writing/tool-calls-need-contracts Mon, 20 Jul 2026 00:00:00 GMT A contract-first method for making model-selected tools narrow, observable, reversible, and safe to evolve. AI Architecture Agents https://sohanjagadeesh.com/writing/structured-outputs-still-need-validation https://sohanjagadeesh.com/writing/structured-outputs-still-need-validation Mon, 20 Jul 2026 00:00:00 GMT How to turn schema-constrained model output into a trustworthy application boundary instead of a well-formed assumption. AI Architecture Reliability https://sohanjagadeesh.com/writing/protect-retrieval-from-prompt-injection https://sohanjagadeesh.com/writing/protect-retrieval-from-prompt-injection Mon, 20 Jul 2026 00:00:00 GMT A defensive architecture for treating retrieved documents as evidence without letting them become hidden instructions. AI Security Retrieval https://sohanjagadeesh.com/writing/evidence-belongs-in-the-interface https://sohanjagadeesh.com/writing/evidence-belongs-in-the-interface Mon, 20 Jul 2026 00:00:00 GMT A portfolio build log about turning technical claims into visible artefacts, decisions, and inspection paths. AI Design Software https://sohanjagadeesh.com/writing/a-comparison-needs-a-stop-rule https://sohanjagadeesh.com/writing/a-comparison-needs-a-stop-rule Sat, 16 May 2026 00:00:00 GMT A small companion note to the model comparison framework about deciding when enough evidence is enough. AI Models Evaluation https://sohanjagadeesh.com/writing/keep-the-corpus-visible https://sohanjagadeesh.com/writing/keep-the-corpus-visible Sun, 08 Mar 2026 00:00:00 GMT A retrieval build log about making source coverage and document state inspectable. AI Retrieval Software https://sohanjagadeesh.com/writing/citations-are-part-of-the-interface https://sohanjagadeesh.com/writing/citations-are-part-of-the-interface Tue, 27 Jan 2026 00:00:00 GMT A note on designing citations for inspection, not merely displaying a source marker. AI Design Evaluation https://sohanjagadeesh.com/writing/a-smaller-context-can-be-a-feature https://sohanjagadeesh.com/writing/a-smaller-context-can-be-a-feature Wed, 19 Nov 2025 00:00:00 GMT Why giving a model more text is not automatically the same as giving it better context. AI Evaluation Workflow https://sohanjagadeesh.com/writing/good-defaults-are-documentation https://sohanjagadeesh.com/writing/good-defaults-are-documentation Fri, 05 Sep 2025 00:00:00 GMT A design note on how a sensible starting state can explain an interface before help text does. Design Software Workflow https://sohanjagadeesh.com/writing/chunking-without-a-magic-number https://sohanjagadeesh.com/writing/chunking-without-a-magic-number Mon, 28 Jul 2025 00:00:00 GMT A short note on treating chunk size as a retrieval decision instead of a copied constant. AI Retrieval Evaluation https://sohanjagadeesh.com/writing/what-this-site-is-becoming https://sohanjagadeesh.com/writing/what-this-site-is-becoming Wed, 14 May 2025 00:00:00 GMT A build log for turning a finished-looking portfolio into a place that can keep changing. Design Workflow Web https://sohanjagadeesh.com/writing/retrieval-is-a-ranking-problem-first https://sohanjagadeesh.com/writing/retrieval-is-a-ranking-problem-first Sat, 22 Mar 2025 00:00:00 GMT Why answer quality often depends on the ordering and cutoff of evidence before a language model begins to write. AI Retrieval Architecture https://sohanjagadeesh.com/writing/model-comparison-without-fake-certainty https://sohanjagadeesh.com/writing/model-comparison-without-fake-certainty Thu, 09 Jan 2025 00:00:00 GMT A dated, reproducible framework for comparing models on the work that matters, without pretending preferences are universal benchmarks. AI Models Evaluation https://sohanjagadeesh.com/writing/pipeline-before-agent https://sohanjagadeesh.com/writing/pipeline-before-agent Tue, 26 Nov 2024 00:00:00 GMT A decision rule for reaching for agentic loops only after a visible, testable pipeline stops being enough. AI Workflow Architecture https://sohanjagadeesh.com/writing/a-practical-anatomy-of-a-rag-evaluation-set https://sohanjagadeesh.com/writing/a-practical-anatomy-of-a-rag-evaluation-set Tue, 17 Sep 2024 00:00:00 GMT A field guide to building retrieval and generation examples that reveal useful failure modes instead of producing one blended score. AI Evaluation Retrieval https://sohanjagadeesh.com/writing/rag-demos-are-easy-retrieval-systems-are-not https://sohanjagadeesh.com/writing/rag-demos-are-easy-retrieval-systems-are-not Sat, 03 Aug 2024 00:00:00 GMT A practical map of the failure modes that appear after the first convincing retrieval augmented generation demo. AI Retrieval Evaluation