AI Agent Starter Kit

Problem

Most RAG and agent tutorials stop at a happy-path demo. Production needs evaluation, observability, cost control, and patterns you can reuse without rewriting glue code every sprint.

Role

Author and maintainer of an open-source starter kit — building production-oriented AI infrastructure in public, with write-ups and a tutorial series.

What it includes

• Three agent patterns — RAG, tool-calling, and workflow-style agents.
• Evaluation harness — so you can measure answers instead of eyeballing demos.
• Observability — LangFuse-oriented tracing for runs and failures.
• Hallucination-aware checks — guardrails beyond “the model said so.”
• MCP integration — tool access without hard-wiring every API into the agent core.
• Cost-aware habits — embed once, stable document IDs, skip redundant indexing (see the production RAG post).

Stack

TypeScript, LangChain, LangGraph, OpenAI, Pinecone, MCP.

Outcomes

• A public, cloneable baseline instead of one-off notebook glue.
• Paired long-form writing on production RAG so the kit isn’t code without context.
• Clear separation between demo toys and shippable patterns (eval, observe, recover).

Tradeoffs

• Opinionated defaults (LangChain/LangGraph ecosystem) for speed of adoption; teams on other stacks still get the architectural ideas.
• Open-source scope stays focused on starters and harnesses, not a full hosted platform.

Links

• Source: github.com/mamenesia/ai-agent-starter-kit
• Deep dive: Building a Production RAG Pipeline That Doesn’t Hallucinate