← All selected work

AI systems architecture

State-aware AI orchestration across enterprise data

A conversational architecture coordinating memory, attachments, retrieval, specialist tools, and synthesis.

Primary evidenceMulti-agent

production-oriented workflow

Situation

What the system needed to change

Enterprise users needed to query structured tables and uploaded files through one conversational surface without collapsing every responsibility into one prompt.

Constraints

The boundaries mattered.

  • 01

    Context and memory had to remain explicit and inspectable

  • 02

    Retrieval required metadata-aware filtering and deduplication

  • 03

    Binary attachments and tabular data needed different processing paths

Decisions

The important engineering choices

  • Separated orchestration, specialist knowledge tools, and final synthesis
  • Added metadata injection, dynamic filters, and hash-based retrieval controls
  • Designed natural-language access across Databricks tables and uploaded CSV or Excel files
  • Moved suitable orchestration from Databricks to n8n to reduce operating complexity

Outcome

What the work left behind

  • An executable multi-agent workflow rather than an architecture-only exercise
  • Clearer boundaries for context, retrieval, progress, and final response generation
  • A path to lower-cost and more maintainable orchestration

Start with the problem

Have a system with similar constraints?

Send the workflow and the boundary that makes it difficult. The first conversation is about fit, not a generic sales deck.