01 · Why this exists.
Most teams don’t have a knowledge problem. They have a process problem dressed up as a knowledge problem. Decisions get made; nobody writes them down. Commitments get spoken; nobody tracks them. Failures happen; the lesson lives in one person’s head. Onboarding takes a quarter not because the docs are bad, but because the docs aren’t where the work is.
The fix isn’t another search box. It’s a system that watches work happen and makes the structure of it legible, to the team, to new hires, and to AI tools the team is already using.
That system is what Pulse is.
We are not building enterprise search. We are building the structured archive of how a 200-person company actually decides things, and the action layer that follows through.Pulse, internal kickoff doc, March 2024
02 · Things we won’t do.
The list of nevers.
- ✕Send a message you didn’t approve. Every external write, Slack DM, Linear ticket, calendar invite, doc comment, passes through your inbox first. The only auto-actions are the ones you explicitly allowlist. The default policy is
draft → approve. - ✕Expand permissions. If you can’t see a thread in Slack, you can’t see it in Pulse. ACLs are mirrored at retrieval, not filtered after the fact. We will lose deals over this, and we’ll keep losing them.
- ✕Train on your tenant data. Your decisions, commitments, and code don’t go into anyone’s training set. Per-tenant encryption keys; BYOK on Scale. Cross-tenant query is a runtime impossibility, not a policy.
- ✕Hide an answer behind a vague confidence band. Every answer carries a calibrated 0–100% confidence. Below threshold, Pulse refuses to synthesize and returns the closest sources. “I don’t know” is a first-class answer.
- ✕Pretend an inferred sentence is a cited one. Lime underline = grounded in a source. Iris dotted underline = inferred from context. Different colors, different contracts. Never silently mixed.
- ✕Optimize for engagement. No streaks, no notifications you didn’t ask for, no daily dopamine. Pulse is something you check when you’re working, then close. Calm is the feature.
- ✕Add a feature without a kill switch. Every detector, every action class, every surface can be turned off per-tenant. No usage gates, no upsell-by-friction.
03 · Things we always do.
The list of always.
- ✓Cite the source. Every passage, every claim. Hover gets you the doc; click gets you the original. The trust contract is at the sentence, not the vendor.
- ✓Surface the conflict. When sources disagree, Pulse shows both and asks which is current. Older guidance vs newer is a first-class signal, never silently resolved.
- ✓Decay gracefully. Docs lose retrieval rank as they age. Decisions get re-evaluation dates. Skills past 0.6 drift score flip to
DRIFTED. Knowledge that ages should look its age. - ✓Allow undo. Every external write has a 5-minute reversal window. The DM you sent can be recalled. The ticket you created can be deleted. The invite you scheduled can be declined.
- ✓Show you what we know about you. A built-in
/medashboard lists every entity Pulse has extracted with you as subject. Exportable. Deletable for non-shared content. - ✓Work with what you have. 9 connectors live, MCP server next. SKILL.md export today. Bring your own LLM key. Federate without sharing data. Pulse adapts to your stack, not the other way around.
04 · Where we stand in the category.
The category has two ends and an empty middle.
One end is Glean, Guru, and the rest of the enterprise-search class. They’re built for the Fortune 500: 4-week IT-led setup, vendor-level trust, per-seat pricing that scales with headcount. They work, for orgs that can dedicate a quarter to onboarding and a procurement cycle to enterprise contracts. They don’t fit a 50-person engineering team.
The other end is Notion AI, Slack AI, and the bolt-on chat-on-your-docs tools. They’re cheap and quick. They also can’t model a decision, can’t track a commitment, can’t draft an action. They’re a feature, not a substrate.
Pulse is the middle. A process graph for software teams of 5–500. Conversational onboarding, not connector-first. Outcome pricing, not seat pricing. Sentence-level trust, not vendor-level trust. Each one is incompatible with the legacy economics that price out mid-market teams, which is why nobody else is building it from this angle.
The hardest thing about building Pulse isn’t the retrieval, the extraction, or the action layer. It’s the patience to keep saying no to the features that would make us look more like the category, and worse for the people we’re building for.
05 · The bet.
The bet is that structure compounds. A doc graph is a snapshot. A process graph is a record of how a company thinks, and a record of how a company thinks gets more valuable the longer you have it.
The bet is that permission is a feature, not a tax. Teams don’t want a tool that knows everything. They want a tool that knows what each person is allowed to know, and is honest about what it can’t see.
The bet is that calm beats loud. The best tool is one you check three times a day for ninety seconds, and trust the rest of the time to surface the things that matter without paging you.
The bet is that action is the missing layer. Suggestions that someone has to act on are still work. A draft you can approve in one click is a follow-through that would otherwise have fallen on the floor.
If we’re right, Pulse is the operating layer for software teams in the next decade. If we’re wrong, we’ll have built one of the more honest pieces of B2B software in the category. Either way, we’re not building anything else.