Build a multi-step agent with LangGraph in Python that classifies tickets, routes them, and replies. Real code, state management, human-in-the-loop.

Agent that classifies incoming tickets, decides between auto-reply or human escalation, drafts the response, and logs to CRM.

Build a multi-step LangGraph agent for SaaS support

Define the graph State

TypedDict with fields the agent accumulates between nodes: ticket, classification, action, response, conversation_history.

Implement the classification node

Python node that calls Claude Haiku to classify the ticket into categories (billing, technical, general, urgent) and adds the classification to state.

Conditional routing based on classification

Conditional edge that routes to auto-respond for simple tickets or to human-review for urgent or complex tickets.

Response generation node

Call Claude Sonnet with a category-specific system prompt, optional RAG against internal docs, output with citations to the knowledge base.

Human-in-the-loop with interrupt

Set interrupt_before on the send node. When the agent wants to send something sensitive, it pauses and waits for human approval via endpoint.

Persistence and observability

Postgres checkpointer so the graph survives crashes. Structured logging of each state transition for debugging.

LangGraph tutorial: your first multi-step agent for SaaS support automation

Practical RAG implementation with Next.js, Supabase pgvector and OpenAI. Index documents, semantic search, augmented answers. Real code, no abstractions.

Index documents, run semantic search, and assemble LLM-augmented answers using Next.js + Supabase pgvector + OpenAI.

Implement RAG in a Next.js SaaS

Enable pgvector in Supabase and create the embeddings table

Turn on the pgvector extension in your Supabase project and create a table with columns for the text chunk, the embedding (vector 1536), and metadata.

Chunk documents and generate embeddings

Split each document into 400-800 token chunks with overlap, generate embeddings using text-embedding-3-small from OpenAI, and insert into Supabase.

Semantic search with cosine similarity

Create a Postgres RPC that, given a query embedding, returns the top-k most similar chunks using pgvector's <=> operator.

Augment the prompt with retrieved context

In the Next.js handler, embed the user question, retrieve top-5 chunks, and inject them into the system prompt before calling the LLM.

Monitor costs and metrics

Track tokens per request, cache hit rate, and latency. Log the augmented prompt for debugging when an answer is bad.

RAG step by step for your SaaS: from PDF to chatbot in 2 hours

A 4-quadrant framework to decide which processes to automate first in your SaaS. How to score impact vs effort, common mistakes, LATAM examples.

Which processes to automate first in your SaaS (a 4-quadrant framework)

Real n8n vs Make comparison: USD pricing, cost per execution, self-hosted vs cloud, useful integrations for LATAM SaaS, and a decision tree.

n8n vs Make in 2026: which one fits a LATAM startup (with real numbers)

What Model Context Protocol (MCP) is, why it changes AI integration in SaaS, and when to adopt it in your product. No hype, concrete cases.

Model Context Protocol (MCP) explained for SaaS founders

Integrate OpenAI in Next.js without the bill spiraling: model selection, caching, streaming, per-tenant rate limiting, and budget caps.

Wrap the SDK with telemetry, pick the right model per call, cache repeated prompts, stream responses, per-tenant rate limiting, and budget caps.

Integrate the OpenAI API in a Next.js SaaS with cost discipline

Wrap the SDK with telemetry

Centralize all OpenAI calls behind one module that logs model, tokens, latency, tenant_id, and estimated cost per request.

Pick the model dynamically per task type

Route every call to the cheapest model that can do the job: gpt-4o-mini for classification/extraction, gpt-4o for rich generation, o3-mini for reasoning.

Cache repeated prompts in Upstash Redis

Deterministic hash of prompt + parameters as key. Cache hit saves 100% of cost and latency. TTL 7-30 days depending on case.

Stream responses for better UX

Use Server-Sent Events to return tokens as they're generated. UX improves 5x and reduces client-side timeouts.

Per-tenant rate limiting

Limit requests per minute and monthly tokens per tenant. Sliding window in Redis. Return 429 with Retry-After header.

Budget caps with kill switch

Track per-tenant monthly accumulated spend. If it exceeds the cap, return an informative error instead of continuing to charge.

How to integrate the OpenAI API in a Next.js SaaS without blowing up costs

The 7 mistakes I see most in AI projects at LATAM SaaS startups, with real cases and how to avoid each. Consulting usually doesn't write about failures — this one does.

7 common mistakes implementing AI in a startup (and how to avoid them)

Real USD ranges per implementation tier: diagnosis (1.5K-2.5K), agents (5K-15K), product features (10K-30K+). Plus hidden costs and ROI math.

How much does it cost to implement AI in a SaaS startup in 2026

Hourly and project rates in LATAM in 2026. Differences between independent consultant, agency, and in-house ML engineer. What to expect in a quote.

How much does an independent AI consultant charge in LATAM in 2026

The 7 questions you should ask before hiring an AI consultant or agency for your SaaS. How to separate marketing from real technical capability.

How to choose an AI consultant: 7 questions to ask before signing

Real cost comparison between Claude and ChatGPT API for LATAM SaaS in 2026: per-token pricing, Spanish tokenization overhead, quality, and which one wins where.

Claude vs ChatGPT API: how much each response costs in Spanish in 2026

Anonymized case of a LATAM B2B SaaS that automated support classification + first response. Diagnosis, stack, real costs, year-1 ROI.

Case study: how a LATAM SaaS automated support and saved 380 hours/month

Understand the differences between Claude's thinking and non-thinking models, learn when to use each mode, and discover how to optimize performance with thinking budgets and adaptive reasoning.

Claude Thinking Models: When to Use Extended Thinking vs Standard Mode

A structured methodology that integrates AI agents into the software development lifecycle — from PRD to production — keeping humans in strategic roles while agents handle implementation, testing, and documentation.

AI-Driven Development: A Methodology for Teams

Learn how to create and distribute your own plugin marketplace for Claude Code, extending the AI with custom skills, agents, and tools tailored to your workflow.

Create Your Own Plugin Marketplace for Claude Code

✅ My complete methodology for developing software using AI agents, from planning to documentation.

Articles about AI