Production AI
Why Your RAG System Works in a Demo but Fails in Production
A retrieval pipeline can look reliable in development and behave very differently under production conditions.
8 min read
01
Building AI
beyond the demo.
I'm Revat Saharan. I build production-grade AI systems and write about the engineering decisions behind making them reliable.
SELECTED WRITING
Engineering notes from building AI systems that have to work.
SELECTED WORK
Systems built to solve real AI engineering problems.
Evaluation trace12 runs
Retrieval Generation
01 RAG Evaluation System
Evaluating RAG Beyond a Single Score
An evaluation system that separates retrieval and generation failures, making it easier to understand where a RAG pipeline is actually breaking.
Python · RAG · Evaluation
02 LLM Reliability
Failure-Tolerant LLM Pipelines
A production-oriented LLM pipeline implementing validation, retries, fallbacks, timeouts and observability for more predictable application behavior.
LLM · Reliability · Observability
pipeline.run
✓ validate input
↳ schema valid
✓ model request attempt 02
↳ fallback resolved in 482ms
✓ trace complete
System map
Input
Retrieval
Tracing
Deploy
03 AI Systems
Production AI Reference Architecture
A reference implementation combining retrieval, evaluation, tracing, caching and deployment patterns for production AI applications.
Architecture · Infrastructure · AI Engineering
ABOUT
I'm Revat, an AI Software Engineer focused on building AI systems for production.
My work sits at the intersection of software engineering and applied AI — with a focus on reliability, evaluation, retrieval, observability, and the infrastructure required to move AI applications beyond prototypes.
More about me →