Kevin Astuhuaman AI product study | Synthetic data GitHub

Public-safe product reconstruction

AI Investigation Workbench

Original interface, fictional commerce data, and no PayPal systems or assets.

Synthetic scenario No production data
01 Scope 02 Plan 03 Evidence 04 Human review

Before the agent runs

The user sees the scope, sources, and boundaries first.

The agent cannot send alerts, change checkout behavior, or make a production decision. It can assemble evidence and recommend the next accountable action.

May doRead approved synthetic sources
Must showEvidence, contradictions, and unknowns
May not doRoute an escalation without approval

Bounded plan

Five read-only steps, visible before execution.

Awaiting approval
  1. 01
    Compare funnel movementAggregate analytics | Locate the change
    Read
  2. 02
    Inspect behavioral patternsSession events | Explain the struggle
    Read
  3. 03
    Correlate service healthService telemetry | Test a system cause
    Read
  4. 04
    Search similar issuesIssue history | Find prior evidence
    Read
  5. 05
    Estimate impact rangeDerived result | State uncertainty
    Compute

Evidence timeline

What happened and why the user struggled.

4 of 5 checks
Funnel analytics Completion fell 12% after payment selection. Observation | Synthetic source A-104
Behavioral sessions Repeated payment attempts and contradictory messages clustered in one flow. Observation | Synthetic source B-220
Service telemetry One source timed out. No healthy result was inferred. Unknown | Retry available
Issue history A prior issue shares two signals but differs by platform. Contradiction | Synthetic source D-041

Ranked hypothesis

A retry-state defect is the leading explanation.

Human review required
Confidence band Moderate
No decimal precision: one diagnostic source is unavailable.
Observation
Retries increased before abandonment.
Inference
The state transition may be presenting conflicting next actions.
Contradiction
The comparable historical issue affected desktop, not mobile.
Unknown
Service telemetry is incomplete. Confirm before remediation.

Human review

The agent drafts the handoff. A person owns the action.

Not yet routed
Proposed owner Checkout experience
Impact range 1,100-1,500 synthetic sessions
Draft diagnosis

Returning mobile customers encountered conflicting payment retry states. Behavioral evidence supports the pattern; service telemetry remains incomplete.

Routed after approval

The evidence package now has an accountable owner.

No production action was taken. The audit trail records the agent plan, source results, missing evidence, hypothesis, reviewer, and routing decision.

Closed without a claim

Incomplete evidence stays incomplete.

The investigation preserves its evidence and unknowns without converting uncertainty into a confident diagnosis.