Context-grounded responder
The Claude agent receives conversation history, app-specific knowledge and product metadata such as version, device and subscription tier before producing a response.
Autonomous customer operations
SupportAgent
An AI support team that knows exactly when it needs a human.
SupportAgent monitors customer conversations across multiple products, grounds responses in each app's knowledge base and routes every case by confidence, sentiment and policy. Routine questions can be handled immediately while refunds, unknown bugs and sensitive situations enter a native human-review queue with full context.
support intelligence
Live routing queue
How do I restore my purchase?
0.94
AUTO REPLYThe app keeps crashing.
0.42
ESCALATEI need a refund.
POLICY
HUMAN REVIEWKNOWLEDGE BASE
CLAUDE AGENT
MACOS OPERATOR
Private product repository. Architecture facts are derived from the implemented agent, email webhook, realtime macOS client and detailed behavioral specifications.
Multi-app
isolated knowledge bases
<0.5
confidence escalation
6
explicit safety triggers
Realtime
native operator queue
System anatomy
Complexity broken into reliable boundaries.
Structured decision output
Every answer includes confidence, category, sentiment, escalation state and a concise issue summary, turning model output into an operational decision.
Risk-based human handoff
Refunds, low confidence, unknown bugs, user frustration, repeated failure and explicit human requests bypass autonomous resolution.
Learning from corrections
Operator overrides are captured as correction signals and aggregated into suggested knowledge-base improvements for future conversations.
Engineering judgment
The decisions behind the build.
Seniority is not the number of technologies in a project. It is the ability to identify failure modes early and choose boundaries that keep the system understandable after launch.
Autonomy is conditional
The agent earns the right to respond from grounding, confidence and policy checks. Sensitive categories do not become autonomous because the wording sounds certain.
Escalation carries context
The operator receives the complete thread, AI summary and suggested response so human review is a fast continuation rather than a workflow restart.
One system, multiple products
Each app keeps separate knowledge and identity while the operator gets one shared support surface across the complete product portfolio.
Production stack
Chosen for the system.
ClaudeProduction promptingSupabasePostgreSQLEdge FunctionsTypeScriptResendRealtimeSwiftUImacOSVitest