Vataman.org

The internal work-OS for a digital-advertising agency — a DataStore plus Graph system that captures source activity, distills the why behind the work, and hands that context to people and AI agents.

New here? The engineering sources of truth are docs/START-HERE.md (current state + roadmap) and CLAUDE.md (the build guide).

The bet

Agencies run on judgment that lives in people's heads and evaporates across email, chat, ad platforms, and contracts. AI agents are getting cheap; context is the scarce thing. Vataman's wager: the durable moat isn't retrieving context, it's distilling it — turning scattered activity into a structured, reusable record of what happened and why, and sharpening that distillation with human feedback.

So vataman is not another dashboard. It is being split into DataStore (source objects, events, intake, mappers, raw docs/transcripts, Google Ads API sync) and Graph (decisions, edges, observations/learnings/assumptions/conflicts, distillation), exposed to agents over MCP.

How It Works

  • DataStore ingests the agency's data and keeps source records durable and inspectable.
  • Graph consumes stored data and turns it into context: decisions, relationships, observations, learnings, assumptions, and conflicts.
  • The Rails app provides the public site, logged-in collaboration shell, CTO console, and bounded in-app LLM operations through the Llm seam.
  • The Agents live in the tools the operator already uses (ChatGPT, Codex, Hyperagent). They read and write through MCP, while humans confirm or correct the important outputs. Those corrections make the next pass better. That feedback loop is the moat.

The first agent, Sage, enriches each organization and person from their public footprint and distills a SOUL.md — a sharp, sourced profile of who they are and how to work with them. Every claim is cited, nothing is fabricated, and a human approves it before it counts.

Philosophy — Information before Technology

Get the information right — the events, the relationships, the why — before reaching for fashionable technology. Boring primitives (Rails, Postgres, markdown) over complexity. The knowledge graph is the product.

Status

Early and honest: the project is mid-rebuild. The current pass is modularizing the monolith into enforced lib/datastore and lib/graph packages before extracting public-site or collaboration boundaries. For real, current status see docs/START-HERE.md.

Stack

Rails 8 · Postgres · Google Cloud Run · Hotwire / ViewComponent · good_job · MCP. Details in CLAUDE.md.