Ramakrishna Reddy·Aug 26Fail safe, not fail silent: what happens when your safety gate itself crashesParts 1 and 2 of this series built a picture: deterministic layers decide, an LLM proposes within a cage, five gates guard every autonomous…
Ramakrishna Reddy·Aug 26Designing a document-processing pipeline: from webhook to human reviewA user photographs a lab-style document and sends it through a messaging channel. Minutes later they receive a clear, reviewed explanation…
Ramakrishna Reddy·Aug 25A tiny expression DSL beats eval(): safe rule evaluation for agentic systemsPart 1 described a nightly agent whose decisions come from a pure-function evaluator over rules-as-data. This part is about the question…
Ramakrishna Reddy·Aug 25The modular monolith is a system design answer, not a compromiseSay “we run a monolith” in a system design interview and watch the interviewer’s pen pause. Say it at a meetup and someone will gently ask…
Ramakrishna Reddy·Aug 19Constrained JSON output in production: schemas, retries, and the confidence floorEvery LLM feature in this series shares one unglamorous dependency: the model’s answer has to land in code. The classifier from Part 1…
Ramakrishna Reddy·Aug 19Five safety layers for LLM-initiated actions in a regulated domainPart 1 of this series was about an LLM that classifies. This part is about something more dangerous: a system that acts — sends the…
Ramakrishna Reddy·Aug 3Deterministic first, LLM second: a layered triage pipeline where the model never gets the last wordAt 2am, a patient sends a message to a clinic’s chat line: “chest feels tight, should I worry?” Nobody is at the desk. A system has to…