Twill Brain

Analytics Context <> LLM

2 min read

Analytics Context <> LLM

Date: 2025-06-06 · Duration: 99.58999633789062 min Organizer: Nader Attendees: Ahmad, Martin, Nader

Summary

  • Ahmadreza presented a comprehensive overview of the current data infrastructure, detailing the extraction of ~100,000 transaction lines from 5 database tables over the past 6 months.
  • The existing data processing pipeline requires significant time, and the EDA JSON output currently reaches ~5k lines, which may hinder real-time data integration for user logins.
  • Daily ingestion trigger was agreed upon as optimal for KPI calculations due to batch settlement constraints, while user login triggers were dismissed due to data staleness risks.
  • Proposed storing KPI summaries in Postgres DB, structured to include recent and historical comparisons, ensuring compatibility with existing onboarding JSON schema.
  • Current EDA JSON is too large for LLM prompts, necessitating context size optimization strategies for improved efficiency in usage.
  • N8N-based MCP tools showcased successful integrations with Stripe, Mercury Bank, and Pay Engine, facilitating direct API interactions without requiring database ingestion for rapid demos.
  • The meeting discussed a split in development focus, proposing separate tools for merchant interactions and management functionalities to enhance user access and system organization.
  • The integration of KPI calculations within the existing business data service was favored over a separate analytics service to simplify architecture.
  • A standardized table schema design was proposed for KPI storage, consisting of three essential columns: value, calculated date, and description.
  • Next steps included task assignments to key team members for testing and development of the proposed systems, emphasizing collaborative contributions for optimal implementation.

Action Items

  • Martin Elias Costa Connect one MCP to a tenant and test integration from backend side without UI to enable LLM communication with MCP tools (01:16:15)

Ahmadreza Abdoli Create postgres table for KPI summaries per tenant with automated end-to-end implementation from ingestion trigger (01:17:15) Clean up and optimize JSON output function to be more consumable by LLMs with two different formats - one for SQL agent and one for business advisor agent (01:34:26) Expand KPI list to include gateway data and payment method analytics beyond basic transaction table KPIs (01:35:45)

Michael Wright Follow up on Luqra API key request to enable MCP integration (01:11:15)

Nader Atrchin Discuss with Sparsh about helping with table creation and analytics functionality setup in business data (01:31:47)

Graph View