AI Operating Layer for Founder-Led Sales

Assumptions

Shivani drafts every first-touch and approves every reply. Over time, I and the agents may draft reply candidates for her review — her sign-off is always required.
Founder-led sales is the only motion that counts right now. I build and run the system around Shivani, with an outbound system architect agent as co-operator.
AI is the intelligence layer — workflow orchestration, lead research, content intelligence, performance analytics. It may draft reply candidates on request; it never ships without her approval.
Work happens across surfaces — our OS, Claude Code, Slack, Gmail, Day AI, Apollo, and any others.

Principles

We own the coordination plane. We rent everything else.
Shivani drafts every first-touch; she approves every reply. AI briefs her, cleans prose on demand, and — for replies only and on request — may propose a draft for her review. First-touches are hers from scratch.
AI organizes context, it doesn't prescribe. What they said, who else is at the account, what the connector flagged, what her past writing shows — facts, not directives.
No auto-send. Nothing ships without Shivani's sign-off.
Most outbound stacks end up ops-heavy and learning-light. We flip that ratio.

Optimizing For

Learning quality and speed
Strategy-to-ops ratio
Reducing new experiment setup time

Tool	Role
Day AI	CRM + context substrate
Apollo	sequenced execution + prospect DB
Happenstance	external relationship graph
Autobound	social + job-change signal intel
Gmail	1:1 sends + forward detection
LinkedIn CM	air cover via matched-audience ads
Outbound OS (ours)	coordination plane + app modules — Supabase · Next.js · Vercel AI SDK · OpenRouter · Composio · memory store · stream / tool adapter

01Example modules inside the Outbound OS — spin up as needed +

These are representative modules we stand up on our own infrastructure. We add or retire them without touching the coordination plane — as the motion moves from founder-led to enterprise, the composition shifts; the core loop doesn't.

Mission Control	The writing queue — not the approval queue. Briefs ready for Shivani to write, replies to handle, intro-asks queued, signal opportunities, account decisions for the day. Minimal one-line rows; focused row expands to quote + glance strip + CTA.
Research Dossier	Per-intent structured context shown left-side of the composer: what happened, what they said (quoted verbatim), the thread, account facts, sensitivities as discrete facts, connector's private notes, observations from her own past writing, and a "noticed" strip with provenance. Organized context, no directives.
Composer	The empty editor. Subject + body blank by default; Shivani writes. On-demand assists: Clean up selection · Tighten paragraph · More Shivani · Draft-for-me (paragraph-scoped, opt-in). Send dispatches via Apollo or Gmail. No auto-send. Copy-to-Gmail escape hatch always available.
Research Deck · 12 slides	Per-lead deep research on demand: cover · psychology · career · public voice · company · org map · trust graph · pain signals · competitive · prior engagement · sensitivities · observed play. Vetted across podcasts, filings, Happenstance, /ledger. Cached 7 days, refreshable per slide.
Thread Surface	First-class chronology with each lead — intros with connector private notes, every exchange, drafting-now state. Parallel multi-thread view across all leads at the same account. Deep-linkable.
Pattern Library	Reference library Shivani pulls from when writing. Holds angles she's leaned on, claims that land in replies, voice observations drawn from her own sends, anti-patterns. Updates as new patterns are noticed — never generates drafts.
Account Central	A live plan per target org. Coordinates contacts across the buying committee, picks warm-vs-cold path via Happenstance, orders threads, and schedules next moves. Ranks intro-asks with template messages she can edit.
Social Signal Inbox	A curated daily list of prospect posts worth engaging with. Manual execution by Shivani — no auto-posting, no auto-comments.
Learning Exchange	Classifies replies, calls, and intro outcomes into event types and routes the lessons back into the Pattern Library, positioning, product, and pricing. Feeds the "noticed" strips across the app with provenance.