Anlytic

▌ Receipts

Metric

60–70%

LLM API cost reduction via dual compression

Metric

8

Visualisation types generated end-to-end from plain English

Metric

0

Destructive operations without human approval

Metric

MCP

Runtime tool loading for extensibility

▌ What got built

1. § 01

   Shipped a production multi-agent AI analytics platform that routes user queries through a Claude Haiku classifier to specialised agents — chart generation, dashboard editing, data management, and general analytics — built on the Vercel AI SDK and Anthropic Claude.

2. § 02

   Enabled non-technical users to build charts in plain English across 8 visualisation types (bar, line, pie, heatmap, icicle, pivot, headline, table). The AI generates the full config and data query end-to-end, removing manual dashboarding entirely.

3. § 03

   Reduced LLM API costs by 60–70% via a dual compression system: stripping tool schema descriptions before sending to Claude, and summarising old tool-call results in message history.

4. § 04

   Eliminated risk of accidental data loss through a human-in-the-loop approval layer — the AI proposes schema changes, row edits, and table deletions, all gated on user sign-off before any destructive operation runs.

5. § 05

   Made the general agent extensible by integrating MCP (Model Context Protocol) for runtime loading of documentation tools, with timeout protection and graceful fallback.

6. § 06

   Kept long chat sessions stable by implementing conversation summarisation via Anthropic's native compaction API — compresses history while preserving critical IDs and user context.