Complexity before execution
Every question is rated from lookup to research-grade before steps are selected; plan size is capped by level.
Governed analytics
EDP Assist
A governed AI analyst that builds the answer live.
Ask a data question in plain English. Review the plan. Watch the analysis build cell by cell in Jupyter. Finish with a cited answer.
30 agents38 skillsRead-only data accessPlan-driven
The problem
Natural-language analytics is easy to demo. Governed, inspectable analysis is harder.
A chatbot can produce a plausible chart in seconds. That does not tell the user whether the question was framed correctly, whether the method changed halfway through, whether the data was altered, or which executed result supports the conclusion.
EDP Assist makes the analytical path visible. It rates complexity first, hard-caps the plan, gates work that deserves approval, narrates every section and closes with an answer that cites executed cells.
Product surface
Built around the loop.
The framework is designed so important guarantees are enforced by hooks, gates and the data library instead of left as prompt-level promises.
Plan-driven multi-agent runtime
One orchestrator composes the plan and one-purpose agents execute narrowly scoped analytical sections.
Approval where it matters
Complex analysis, model training and library installs pause. Simple questions can move immediately.
Narrated Jupyter work
Every section states What, Why and How, runs the analysis, and records a plain-English Finding.
One read-only data door
The guarded data interface accepts read-only analysis paths and refuses write requests.
Cited Answer cards
The final conclusion points back to the executed notebook sections that support each claim.
How it works
Ask. Plan. Approve. Watch. Answer.
01
Ask
Your words remain the canonical analysis goal.
02
Plan
Complexity is scored and a bounded plan DAG is composed.
03
Approve
Only work with meaningful risk or cost pauses for consent.
04
Watch
The notebook builds as a readable analytical narrative.
05
Answer
A cited Answer card closes the question, names limitations and suggests grounded next exploration.
retention-question.ipynbread-only / pass 1 of 2
WHATMeasure activation retention by invite behavior.
WHYTest whether the observed retention assumption is supported by user behavior.
HOWBuild cohorts, compare day-30 retention, estimate uncertainty.
ANSWEREarly teammate invitation is associated with higher day-30 retention in the observed cohort.Cells 5, 8, 11 · association, not causal proof
EDP Assist
A governed AI analyst that builds the answer live.
Questions about access, product collaboration or the underlying system?