GraphSense
Insight discovery that runs 3,412 candidate analyses and shows you eighteen.
Overview
The product is a filter, not a generator. It fingerprints the columns in your connected sheets, builds a knowledge graph, walks it to produce candidate analyses, then removes 99.5% of them across seven named layers. Each surviving finding carries a five-dimension interestingness score and cites the table it came from. The value is not the analyses it ran — it is the 3,394 it refused to show you.
Specifications
USE CASE
Automated insight discovery across connected operational data
WORKFLOW
Fingerprint, graph, generate, filter, rank
OUTPUT
Eighteen findings with source citation and interestingness score
SECURITY
Runs entirely on-premise — no production data leaves the site
Features
- 01Animated seven-layer funnel with per-layer drop counts and what each removes
- 02Five-dimension interestingness on every finding — surprise, magnitude, variance, stability, coverage
- 03Column fingerprinting that infers types from values, not headers
- 04Knowledge graph the candidate generator actually walks
- 05Natural-language questions answered from your own tables, with row counts
Architecture
GRAPHSENSE PIPELINE
────────────────────────────────────────────
┌────────────────────────────────┐
│ Sheets, CSV, XLSX, ERP exports │
└────────────────────────────────┘
↓
╭┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈╮
┊ Column fingerprinting ┊
╰┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈╯
↓
┌─────────────────┐
│ Knowledge graph │
└─────────────────┘
↓
╭┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈╮
┊ Candidate generation — 3,412 ┊
╰┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈╯
↓
╭┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈╮
┊ Seven-layer filter ┊
╰┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈╯
↓
┌─────────────┐ ┌────────┐ ┌─────┐
│ Discoveries │ │ Alerts │ │ Ask │
└─────────────┘ └────────┘ └─────┘
╭┈┈╮ dashed = model step ┌──┐ deterministic stepWhat it will not do
Every screen carries one honest failure record with its recovery path. This is the one buyers remember. A run surfaced 'output fell 18% in week 28' as a top finding. Week 28 was the annual shutdown, and the known-cause layer had no calendar of planned stoppages. The plant calendar is now a first-class input and nine similar findings were retired in the same pass. A second run refused to phrase an operator-variance finding as a performance issue — both operators had the highest output on the cell and the SOP was the outlier.
