← Graph

Building LLM powered applications in Ruby

talk 25 connections

Andrei Bondarev's wroclove.rb 2024 single-speaker talk. Frames generative AI as about to become a standard part of every tech stack (alongside databases, caches, queues, storage) per the a16z vision, with developers building on top of AI systems. Covers: what LLMs are and what they excel at (structuring unstructured data, summarization, classification, translation, content generation, Q&A), the shift from business logic in fat models/service objects to business logic in prompts, AI agents as autonomous LLM-powered programs using tools via function calling, and the reliability/generality trade-off (focused narrow agents are POC-ready; general-purpose reliable agents would equal AGI). Discusses slow adoption (fast-changing field, IP/copyright ambiguity, lack of tooling, risk — Air Canada chatbot lawsuit, GM chatbot agreeing to sell a car for $1), calls prompt engineering 'prompt alchemy', covers jailbreaking (Anthropic's many-shot jailbreaking paper) and hallucinations (GPT-4 trained up to April 2023). Introduces RAG: embed user query, similarity-search a vector store of proprietary docs, merge context into prompt. Explains vector embeddings (OpenAI's Ada model is 1536 dimensions), similarity metrics (Manhattan, Euclidean, cosine), evaluations via human up/down votes, LLM-as-critic using GPT-4, and the RAGAS method (faithfulness, context relevance, answer relevance). Argues against open chat bots due to prompt-injection risk — advocates closed control-panel UIs with narrowly-scoped agents. Live-codes a 'Nerds and Threads' e-commerce AI assistant using langchainrb with six tools (customer management, email service, inventory management, order management, payment gateway, shipping service) backed by SQLite, demonstrating order placement, returns, and inventory updates — each driven by a system prompt describing standard operating procedures. Closes on why Ruby (pragmatism, OOP principles, Sandi Metz's POODR, ability to port Python libs via ChatGPT) and open-source maintenance lessons (be responsive, friendly, helpful). Q&A explores generating executable Ruby code once vs. instructing the LLM per request.

type
talk
talk Building LLM powered applications in Ruby
about
langchainrb tool
Talk centres on langchainrb as the Ruby solution for LLM-powered apps.
talk Building LLM powered applications in Ruby
about
Generative AI concept
Introduces generative AI as the context for Ruby applications.
talk Building LLM powered applications in Ruby
about
Large Language Models concept
Explains LLMs and their strengths.
talk Building LLM powered applications in Ruby
about
Transformers concept
Names Transformers as the underlying LLM architecture.
talk Building LLM powered applications in Ruby
about
AI Agent concept
Covers AI agents, function calling, and the focus/reliability trade-off.
talk Building LLM powered applications in Ruby
about
Retrieval Augmented Generation concept
Explains naive RAG and advanced multi-index strategies.
talk Building LLM powered applications in Ruby
about
Vector Embeddings concept
Describes embedding models and 1536-dimensional vector spaces.
talk Building LLM powered applications in Ruby
about
Vector Database concept
Vector DBs used as the similarity-search substrate for RAG.
talk Building LLM powered applications in Ruby
about
RAGAS concept
Introduced as a quantitative way to evaluate RAG systems.
talk Building LLM powered applications in Ruby
about
Prompt Alchemy concept
Andrei renames 'prompt engineering' to 'prompt alchemy' and argues it's not engineering.
talk Building LLM powered applications in Ruby
about
Jailbreaking concept
Covers jailbreaking techniques including many-shot jailbreaking.
talk Building LLM powered applications in Ruby
about
Hallucinations concept
Discusses hallucinations and knowledge cut-offs as motivation for RAG.
talk Building LLM powered applications in Ruby
about
Nerds and Threads project
Live-coding demo central to the talk.
talk Building LLM powered applications in Ruby
about
Attention Is All You Need resource
Names the 2017 Google paper that kicked off modern LLMs.
talk Building LLM powered applications in Ruby
about
Many-shot Jailbreaking Paper resource
References Anthropic's many-shot jailbreaking paper.
talk Building LLM powered applications in Ruby
about
Andreessen Horowitz company
Cites a16z's vision of generative AI as a core tech-stack component.
talk Building LLM powered applications in Ruby
about
Air Canada company
Air Canada chatbot lawsuit used as a cautionary example.
talk Building LLM powered applications in Ruby
about
General Motors company
GM chatbot agreeing to sell a car for $1 used as a cautionary example.
question Generate code once vs instruct LLM per request
asked_at
Building LLM powered applications in Ruby talk
Audience question during the Q&A.
person Andrei Bondarev
authored
Building LLM powered applications in Ruby talk
Andrei delivered this single-speaker talk.
takeaway Prompts as Business Logic
from_talk
Building LLM powered applications in Ruby talk
Takeaway drawn from Andrei's proposed shift of business logic into prompts.
takeaway Narrow Agent Responsibilities For Reliability
from_talk
Building LLM powered applications in Ruby talk
Andrei's central argument about agent reliability.
takeaway Port Python Libraries With ChatGPT
from_talk
Building LLM powered applications in Ruby talk
Andrei's practical recommendation to Ruby developers missing Python libraries.
takeaway Open Source Maintainer Lessons
from_talk
Building LLM powered applications in Ruby talk
Lessons Andrei shared from running langchainrb as open source.
talk Building LLM powered applications in Ruby
presented_at
wroclove.rb 2024 event
Talk given at wroclove.rb 2024 in Wrocław.

Provenance

Created
2026-04-17 16:17 seed
Last updated in
Building LLM-Powered Applications in Ruby — Andrei Bondar... 2026-04-17 23:20
Read by
77 extractions