Rishi RaiDesigning Reliable Retrieval-Augmented Generation Systems
Technical guide
An engineering guide to retrieval, reranking, citations, evaluation, access control, and conservative failure handling.
Notes and articles
Technical guides by Rishi Rai on reliable RAG, LLM evaluation, AI agents, security, observability, cost, latency, and software engineering for AI.
Evaluating LLM Applications Beyond Accuracy
Technical guide
A repeatable approach to datasets, deterministic checks, rubric graders, pairwise comparisons, human review, and release gates.
Building Production-Ready AI Agents
Technical guide
Control agent behavior with durable state, narrow tools, policy checks, approvals, idempotency, observability, and recovery.
Reducing Hallucinations in Enterprise AI Systems
Engineering analysis
Layer task boundaries, evidence retrieval, abstention, citations, validation, and human review around model output.
LLM Observability, Tracing, and Debugging
Technical guide
Connect model calls to retrieval, tools, policy, validation, token use, cost, privacy controls, and reproducible replay.
Designing AI Automation Workflows That Fail Safely
Technical guide
Use deterministic orchestration, explicit state, idempotency, approvals, compensation, audit records, and bounded retries.
Prompt Versioning and Regression Testing
Technical guide
Treat prompts, models, schemas, retrieval rules, and evaluation data as one versioned and testable release artifact.
Security Threat Modeling for LLM Applications
Security guide
Map trust boundaries and contain prompt injection, tool misuse, retrieval poisoning, data leakage, and resource abuse.
Reducing Cost and Latency in Generative AI Systems
Engineering analysis
Optimize the complete request path through measurement, model routing, context budgets, caching, batching, streaming, and bounded fallback.
Combining Software Engineering Discipline with AI Application Development
Engineering analysis
Build deterministic boundaries, versioned data and prompts, layered tests, observability, secure tool use, and reversible delivery around AI.