An agentic development workflow, with memory

We deployed a dedicated agentic foundation on the client's server: an orchestrator scheduling recurring 24/7 jobs, a persistent long-term memory (dedicated MCP server), a knowledge graph of the ecosystem, reusable skills, and supervised autonomous sub-agents. Sourced RAG (chunking + embeddings + retrieval), completed by a graph + SQL layer. Rule: no source, no answer.

  • Orchestration (Dagster), long-term memory (MCP)
  • RAG: embeddings + knowledge graph + SQL
  • Extractor / Data Analyst / Reviewer sub-agent trio
  • OAuth2 read-only incremental connectors: Gmail, Drive, Odoo, TrackTik, Notion
  • Git versioning + auto-push, no-auto-send email guardrail

Tech Stack

Claude Code, Dagster, MCP memory, RAG (embeddings + graph + SQL), OAuth2, Git

The challenge

Have AI agents work on real business tasks, continuously, with no context loss, no hallucination and never crossing a guardrail. A raw LLM answers from memory and invents; we needed a system that knows, cites its sources and stays under human control.

Result

An on-demand, multi-company data scientist and admin assistant that knows the context, cites its sources and stays accountable to a human.

More projects

A member-care agent, grounded in data

Member requests answered fast and accurately, with drafts in the brand voice and sending left to humans.

An agentic second brain wired to the business

Plain-language questions answered on real business data, with member reply drafts under human control.

End-to-end finance automation: AGICAP, Exact Online and Power BI

A full finance back-office automation: supplier invoices captured, auto-encoded, synced to the accounting system, and reported live, with the accountant only validating.

Agentic file-server migration to SharePoint

A multi-terabyte file server migrated, sorted and tagged, Copilot-ready.

An agentic SEO and GEO visibility loop

A weekly agentic loop that monitors SEO and generative-engine visibility, and proposes prioritised actions.