ombharatiya/ai-system-design-guide: a Markdown reference for AI system design interviews and production work

What ombharatiya/ai-system-design-guide is, and what it is not

This is a documentation repository. The top level contains directories numbered 00 through 19, plus standalone files such as GLOSSARY.md, PATTERNS.md, COURSES.md, TRANSITION_GUIDE.md and RESEARCH-RADAR.md. There is no package manifest visible in the repository layout, no build file, and no source directory in the listing. The README describes the project as "a practical, continuously updated guide to AI system design, RAG architectures, LLM engineering, agentic AI, MCP and A2A protocols, and AI engineering interview preparation." That description, not a feature list, is the honest scope.

The problem it addresses is a reading problem. An engineer preparing for an AI system design interview, or moving from a general backend role into production LLM work, has to assemble a mental model of chunking, vector search, reranking, tool use, evaluation and cost control from scattered blog posts. This repository puts those topics in one tree with a navigation table that maps an intent ("Build production RAG", "Design multi-tenant AI", "Route across models") to an ordered reading path. The audience is the candidate or the working engineer who wants structure, not the team that wants a library to import.

The README is explicit that the online reader at aidaddy.tech is the intended reading surface, offering "instant search, linked chapters, and a cleaner reader". GitHub remains the canonical source, but the README points readers to the site for navigation. If you judge documentation projects by whether you can clone and run something, this one will disappoint you immediately.

How the guide is organised: numbered tracks and cross-linked chapters

The mechanism is a numbered directory tree. 00-interview-prep holds the question bank and answer frameworks. 01-foundations covers LLM internals. 02-model-landscape covers model taxonomy and pricing. 03-training-and-adaptation, 04-inference-optimization and 05-prompting-and-context follow. 06-retrieval-systems is the largest visible track, running from RAG fundamentals through chunking, vector databases, reranking, contextual retrieval, late-interaction ColBERT, multimodal RAG and production RAG at scale. 07-agentic-systems covers agents, tool use and MCP, and loop engineering. The remaining directories run through memory and state, frameworks and tools, document processing, infrastructure and MLOps, security and access, reliability and safety, evaluation and observability, design patterns, case studies, tool-use and computer agents, voice and audio agents, and multimodal generation.

The README's navigation table is the real interface. Each row pairs a goal with a sequence of file paths, so the reader is not left to guess the order. For example, the production RAG path is chunking, then vector databases, then reranking, then production RAG at scale. The advanced retrieval path layers contextual retrieval, ColBERT and multimodal RAG on top. Two standalone eval guides sit at the repository root rather than inside 14-evaluation-and-observability: ai_evals_comprehensive_study_guide.md, described in the navigation as covering Phoenix and Langfuse, and ai_evals_complete_guide_langwatch_langfuse.md, covering LangWatch and Langfuse. That split is a small structural inconsistency worth knowing before you go looking for evals inside the numbered tree.

Several chapters are framed around churn rather than around a single stack. 09-frameworks-and-tools includes a chapter on navigating framework churn, and the README labels it as covering "stale tutorials, version pinning, what to actually learn". That is an admission that the surrounding chapters age, and it is the most useful structural signal in the repository: the guide expects its own framework-specific content to go stale.

Reading it locally: open a chapter from the numbered tree

The README does not document an install command, a package, or a build step, because there is nothing to install. The practical first use is to read a chapter as plain Markdown, either on GitHub or at the online reader the README links to. The README points new readers at the 128-question interview bank, the RAG fundamentals chapter, or the model taxonomy chapter as entry points.

The repository layout lists the chapters as files under numbered directories. The interview bank the README links as the first stop for interview preparation is:

00-interview-prep/01-question-bank.md

The two paths the README lists for learning AI systems fast and for building production RAG are:

01-foundations/01-llm-internals.md
06-retrieval-systems/01-rag-fundamentals.md
06-retrieval-systems/02-chunking-strategies.md
06-retrieval-systems/04-vector-databases.md
06-retrieval-systems/06-reranking-strategies.md
06-retrieval-systems/14-production-rag-at-scale.md

What you should see is prose and diagrams in Markdown, not executable examples. The repository layout shows no test suite and no example application, so do not expect to run a RAG pipeline after opening these files. If you want the searchable reader instead, the README directs you to aidaddy.tech.

Where the guide stops: no code, no versions, no releases

The limitation is structural. This is a prose reference, and the README describes it as a guide, not a toolkit. Nothing in the repository layout suggests runnable reference implementations, and the README does not claim any. If your team needs a working retrieval pipeline to adapt, this repository gives you the vocabulary and the decision criteria, and you write the code yourself.

The second limitation is staleness. The guide covers model pricing, model taxonomy, framework orchestration and benchmark interpretation, all of which move faster than a documentation repository can. The README calls the guide "continuously updated", and the repository lists no releases, so there is no version you can cite or pin. A chapter that names a model tier or a price is only as good as its last edit. The guide partly acknowledges this with its chapter on framework churn, which discusses version pinning and stale tutorials, but acknowledging the problem does not solve it for the reader who is skimming a pricing table.

Third, the repository mixes two audiences. The interview-preparation material and the production-engineering material overlap but do not have the same success criteria. An interview answer framework rewards structured trade-off discussion. A production chapter on FinOps or durable execution rewards specifics. The navigation table does separate the paths, which helps, but a reader who starts at the question bank and reads straight through will hit production chapters written at a different level of detail.

Finally, the licensing is permissive but the content is opinionated. MIT covers the repository, and the README does not describe any separate content licence or attribution requirement for the chapters themselves.

Alternatives: prose guide versus runnable framework

The closest thing to a competing resource in the same space is a full framework such as LangChain, which the guide itself covers in 09-frameworks-and-tools. The difference in approach is fundamental. LangChain ships code: abstractions, integrations and a runtime you install and call. This repository ships explanation: why chunking choices change retrieval quality, how reranking fits after vector search, what a gateway does for fallback and rate limits. If you adopt LangChain you get working primitives and you inherit its release cadence and breaking changes. If you read this guide you get no primitives at all, but you also inherit no dependency.

A second comparison is against a single-author book on AI engineering. A book is edited, sequenced and stable, and it goes out of date as a unit. This repository is a directory of Markdown files that can be edited independently, which makes partial updates cheap and consistency harder. The README's own navigation table, which points at specific file paths, is evidence of that: the structure is maintained by hand, and a renamed file breaks the path.

The honest framing is that these are complements, not substitutes. The guide's chapter on AI gateways and model routing discusses LiteLLM and fallback behaviour; that chapter is useful precisely because a framework like LiteLLM is what you would install. Use the guide to decide what to build, then pick the library.

Editorial conclusion

Adopt it if you are preparing for an AI system design interview, or if you want a chapter-by-chapter reading list for RAG, agent loops, gateways and evals and you are comfortable reading Markdown on GitHub. Do not adopt it if you need runnable code, a reference implementation, or a versioned dependency you can pin. Before you rely on a chapter, open the file on GitHub and check whether the models, prices and framework versions it names are still current, because the repository ships no releases and the README does not describe a changelog or deprecation policy.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. ombharatiya/ai-system-design-guide on GitHub
4. Project website
5. README