Skip to content
All experiments

Agentic AI

A hands-on journey from LLM basics to production multi-agent systems — LangChain, LangGraph, MCP, RAG, guardrails, evals and deployment, built from scratch and measured.

In progressStill being tested

Hypothesis

Learn agentic AI hands-on, not by watching tutorials: build and actually run each stage from raw LangGraph state machines up to a deployed multi-agent system, and record real captured output — including the failures — rather than illustrative examples.

Outcome

17 of 22 modules built and verified, each with a committed, timestamped transcript (commit SHA + UTC time) from actually running the code against live APIs. Modules 20 and 22 aren't built — they need paid Pinecone/Neo4j credentials, not more time. Module 3 (legacy LangChain) was deliberately left unrun, flagged as likely bit-rotten rather than silently claimed working.

Headline evidence: a 31-case eval suite passing 116/117 checks (99.1%), up from 88.6% before three real guardrail bugs it found were fixed — including that "Ignore all previous instructions" was not being blocked, the most common injection phrasing there is. A budget-cap test shows refused requests costing zero model calls. A full pass of every built module costs about $0.09 on a mini model (Groq's free tier covers it entirely).

Explicitly not verified, and said so rather than implied otherwise: nothing is deployed to the cloud — every container runs locally. Both CI workflows have never actually run (they sit in module folders instead of the repo root GitHub reads). The Streamlit UI boots but has never been clicked through. Zero OpenAI or DeepSeek calls were made — provider wiring is verified only up to "fails correctly on a missing key."

AIAGENTSRAG