Google AI Studio
Agent behaviour prototyping and pressure testing
Inbox Pilot · Applied AI
Inbox Pilot is a read-only AI attention layer for Gmail that identifies actions, deadlines, meetings and important decisions while filtering routine inbox noise.
Synthetic or anonymized content only. No real Gmail content will appear here.
01 · The problem
Email systems are optimized for delivery, not attention. Newsletters, receipts, interview invitations, invoice deadlines and security alerts all compete in essentially the same queue.
02 · The product idea
“Does this actually need the user?”
Filter noise → surface what matters → prepare the next step → let the human decide.
03 · First failure / design lesson
During the first prototype, an email asked: “Could we meet Tuesday at 14:00?” The first agent version drafted: “Tuesday at 14:00 works great for me.”
That was rejected because the AI had made a human decision.
04 · System
05 · Guardrails
06 · Engineering proof
No production SaaS deployment is claimed.
Agent behaviour prototyping and pressure testing
Structured email triage
Engineering execution and regression work
Gmail + Calendar read-only integration
07 · Benchmark
synthetic emails
6surfaced
70%suppressed
real inbox emails
3surfaced
93%suppressed from the primary queue
Current state: human validation in progress. These are triage and suppression results, not an accuracy claim.
08 · Privacy
Real Gmail content is processed for triage but should never appear in public portfolio assets. Public screenshots and video use synthetic, anonymized or safely obscured data.
NorthScope AI · Applied systems