JARVIS Core

Context

JARVIS Core started from a simple observation: producing quality content regularly (articles, product copy, posts) takes a lot of back-and-forth — research, drafting, review, formatting. The goal was to build an agent system able to carry most of that chain, while keeping a human in the loop at the decision points.

Problem

“One-shot” content generation tools produce plausible but generic text, with no memory of brand context and no fact-checking. The system needed to plan a content task into sub-steps, fetch verifiable information, and self-correct before proposing a deliverable.

Architecture

The system is organised as specialised agents orchestrated by a state graph (LangGraph):

• a planner agent that breaks the brief into content steps and goals;
• a researcher agent, connected to a RAG knowledge base, that gathers sources and facts;
• a writer agent that produces a structured draft from the plan and sources;
• a reviewer agent that checks tone, consistency and accuracy, sending the draft back for revision when needed;
• a human validation step before publishing, exposed through a FastAPI endpoint and a light dashboard.

Stack

FastAPI for orchestration and the API, LangGraph/LangChain for the agents, a vector store for RAG, and a task queue to parallelise research without blocking the API.

Results

• Shorter content production cycle thanks to parallel research/writing.
• Lower human-revision rate once the knowledge base was enriched.
• [Quantified metrics to add after the next measurement pass]

Screenshots

Screenshots coming soon.