For Data TeamsThe questions your
The questions your
backlog can't get to.
Your queue of "can you look at X?" is forty deep. Every one is pattern-finding in tables you already modeled. DecisionBox is a junior analyst that works unattended, writes its own SQL, and leaves a full audit trail — for the questions even your team wouldn't have thought to ask.
CRITICALSub-hourly temporal
Network-wide SMS surged 10.2× and calls 23.4× in a 10-minute window at NYE midnight — invisible at hourly granularity, a sharp 30-minute spike
SMS peak 360,131 at 23:10 · 12.9× same-time baseline
Telco CDR dataset · 10,000 Milan cells
What it answers
Five questions you keep postponing
01
What patterns exist across two tables I have not joined yet — and are any of them strong enough to warrant building?
The agent hypothesizes joins, runs them, and ranks results by signal. The exploratory work your team defers.
02
What changed in last week’s data that nobody has asked me about yet?
Distribution shifts, null-rate changes, segment regressions — surfaced from the data itself, not from a dashboard tile.
03
Which of my curated segments have broken under their own weight — stale definitions, zero-activity members, orphaned joins?
The agent inspects your modeled layer's output, not your code. Flags the silent failures.
04
Below what time granularity do interesting patterns first appear?
Hourly smooths out most of what happens. The agent scans finer resolutions and only reports where the pattern survives.
05
Given a domain pack, which findings worth flagging to the business team exist in my data — without them having to ask?
Findings tagged by role. Gives your downstream consumers a starting point instead of a blank Slack thread.
From real runs
From a pattern to a playbook
Every finding the agent produces comes paired with a recommendation — a target segment, an expected impact, a concrete set of actions. Three real pairs below, pinned as the agent surfaced them.
HIGHGeographic mirror
Over the holidays Milan lost 51% of internet traffic while Trentino ski-resort cells gained 109% — an exact mirror
Milan −51% · Trentino +109%
Telco CDR · 3,131 cells flagged
size for the surge →
P1Recommendation
Capacity provisioning
Deploy temporary capacity infrastructure at 646 Trentino ski-resort cells for the winter holiday season, sized to the +109% surge
Expected impact
60–80% reduction in congestion events
Target: 646 ski-resort cells5 concrete actions
HIGHService-mix asymmetry
Voice drops 40.4% from weekday to weekend; internet only drops 12%. The service mix is elastic in one direction only. Business-district cells run at a 4–9× ratio; their weekend voice capacity sits idle
−40.4% voice · −12.0% internet · weekend
Telco CDR · 10,000 Milan cells
reallocate the idle
P3Recommendation
Capacity reallocation
Reallocate weekend voice capacity to data for 35 business-district cells
Expected impact
+15–25% weekend data throughput
Target: 2,119 cells5 concrete actions
HIGHTopology concentration
Top 10% of cells carry 48% of all network traffic. Concentration is identical across urban Milan and rural Trentino — a network-topology property, not a city phenomenon
48% of traffic · 10% of cells
Telco CDR · 16,259 cells, 63 days
densify the hot 9 ~
P1Recommendation
Cell densification
Densify or split the top 9 Milan central cells operating at sustained ceiling (peak-to-average ratio 1.10–1.23)
Expected impact
+30–40% effective capacity headroom
Target: 9 central cells5 concrete actions
Every pair above is from a real DecisionBox run. Source datasets are anonymized for the public site.
The honest list
What it does, and what it doesn't
What it does
- Consumes your modeled layer — dbt, a semantic layer, curated marts. Does not try to replace it.
- Every SQL query auditable. Every finding re-validated against the warehouse before it ships.
- Self-hostable. Your warehouse credentials, your LLM provider, your infrastructure.
What it doesn't
- Replace dbt or your modeling layer. It consumes what your team produces.
- Function as a data-quality monitor. Use a tool built for that job.
- Produce clean findings on raw event logs. Point it at curated marts for signal.
app.decisionbox.io/discoveries
Keep reading