AI Engineering in Python: System Design 101

Nirant Kasliwal (~NirantK)




LLM-based applications are a critical component of modern AI workflows. This is a system design primer for ML beginners wanting to focus more on the concepts than specific tools and models. At the same time, I recommend Python-first tools Open Source tools to help beginners good and recent choices.

We use two main driving examples to help folks reason through the problems: Retrieval Augmented Generation — this is the method used for every "Chat with your Data" solution:

Chat with your Data

  1. RAG System Outline — Best Practices
  2. Improving Ranking
  3. Scaling it Up
  4. Improving Reliability

This talk is accessible, welcoming and at the same time enriching to anyone who'd find an opinionated introduction to AI Engineering useful!


  1. Ability to work with REST requests in Python e.g. some FastAPI/Flask/Django is more than enough
  2. Played with or used ChatGPT-like Interfaces

Content URLs:

PyCon India 2023 Slides are here

I'd given a version of this talk recently at IIT Madras' AI4Bharat as well, which was more suitable for academia than PyCon. The talk was very well received with multiple Professors from across multiple departments sitting through the entire session and asking good questions.

Speaker Info:

Me in 3 bullet points:

  1. ACL 2020 NLP Paper on Hinglish, first Hindi-LM: Hindi2Vec
  2. NLP Book with 5000+ copies sold
  3. For Stanford CS230 Deep Learning, Dr. Andrew Ng recommends my work [Awesome NLP](

Spoke at PyCon2019 in Chennai:

As a Machine Learning Engineer, I have:

  1. Deployed Sentence Transformers and Annoy (vector search library) for cosine Similarity powered search in 2018 in production
  2. Managed a team of 3 engineers to build a support chatbot for 1M chat messages per month
  3. Created Hinglish LM Dataset and Model for Hindi-English Code-Switching

As an AI Engineer, I have

  1. Built and deployed Question Answering systems for 3+ years, including production-grade projects with OpenAI LLMs e.g. text-davinci-003, GPT3.5 and GPT4
  2. Built guardrails, evals to reduce hallucinations and deployed summarization and question answering systems

Speaker Links:

Section: Data Science, AI & ML
Type: Talks
Target Audience: Beginner
Last Updated: