The company brain · YC Request for Startups

The models are ready. Your company's knowledge isn't.

That's the real blocker to AI automation now — not the models, but how your team actually works, scattered across docs, code and people's heads. RampBrain compiles it into governed skills an AI agent runs — refunds, incidents, approvals, access — every step sourced, gated on your approval, and sharper every time it's used.

Free plan forever No credit card Live now
RampBrain · refund-policy

Compiled skill · grounded in your sources

Live
refund-policy.md #support-runbook Stripe docs interview: Raj approvals.md
Compiled into a governed decision
Annual plan, cancelled after 2 months → pro-rata partial refund refund-policy.md
10 unused months = $533.33 §4
Stop: $640 > $500 auto-limit — needs Head of Support §7
Partial refund — $533.33 Proposed awaiting your approval
Learns from how it's resolved — and gets sharper
60%
of how a team operates is undocumented — it lives in people's heads
23 hrs
per week lost finding answers and re-doing work the team already knows
$50k
average cost of lost know-how when a senior person leaves
Gated
every agent action waits for one-click human approval — nothing auto-runs

First three figures are widely-cited industry estimates (knowledge-management research).

Plugs into the tools your knowledge already lives in

GitHub Slack Jira Confluence Notion Linear PDF & Markdown 5-min interviews
The Problem

The work stalls on what's in people's heads.

Your team already knows how to handle it. But every decision routes through the one person who remembers the rule — and AI that only answers can't pick up the slack.

💬

"Ask Raj, he knows."

The refund edge case, the deploy quirk, who signs off on what — it lives in one person's head. When they're out, the work waits.

💰

"Is this within policy?"

Every borderline refund, discount, or access request pings the one person who knows the rule. The policy exists — in a doc nobody opens — so the same call gets re-made by hand, all day.

🔥

"The runbook's out of date."

An incident hits at 2am. The runbook was written once and never updated, so on-call guesses — instead of following what actually fixed it last time.

🤖

"The AI can answer. It can't act."

A chatbot surfaces a document and stops. It won't safely run your refund or incident process — because it doesn't know your rules, and nobody trusts it to act without a gate.

Every team has a Raj.

Raj knows which refunds to approve. Raj knows what actually fixed the last outage. Raj knows how things really work around here — and none of it is written down.

Raj is one two-week notice away
from taking all of it with him.

Any policy → a workflow it runs

You don't get three features. You get a compiler.

Point RampBrain at any operating policy and it compiles a governed workflow — with that policy's own decision options and limits. Refunds and incidents ship as showcases; everything else, you compile yourself. No new module, no engineering.

deal-desk-discount.md
your discount policy
deal-desk-discount · skill
decisions: approve · escalate · deny limit: 15%
Discount ≤ 15% on annual plans → approve policy §2
15–25% → escalate to Deal Desk policy §3
! Multi-year terms not documented → escalate, don't guess
Workflows teams compile on day one
Refunds Incident response Discount approvals Access requests Spend sign-off Vendor review Onboarding + any policy you have
How RampBrain works

Not a doc-search tool. A loop that compounds.

Compile how your team operates into skills an agent runs — safely — and that get sharper every time they're used.

01

Ingest & understand

Connect your sources. We parse and structure them into a living map of how your team operates — processes, rules, ownership, decisions.

GitHubDocsSlackInterviews
02

Compile a skill

Any policy becomes a grounded skill — its decision options, limits, steps and guardrails, each traced to a source, and the gaps flagged.

SourcedGap-awareGuardrails
03

Agent proposes, you approve

The agent acts on documented rules only and escalates on a gap. You approve — then it runs the safe step; money and prod changes stay a cited human checklist.

Human-gatedSlack / LinearWebhook
04

Learns from the resolution

Capture how a case was resolved and RampBrain recompiles the skill in place — closing the gap that caused the escalation. Next time, it just handles it.

Self-updatingGap closes
Step 4 feeds step 2 — the brain gets more capable the more your team uses it.
What you get

A real decision — grounded, and gated.

On a live request the agent matches a documented rule, proposes the call — and stops at your guardrail. The irreversible step always stays a human one. Same primitive, any workflow:

Refund decision — Order #4821 · $640 requested
Matched policy rule
Annual plan, cancelled after 2 months → eligible for a pro-rata partial refundrefund-policy.md
Pro-rata on 10 unused months = $533.33refund-policy.md §4
Guardrail
Stop: $640 is over the $500 auto-approve limit — needs Head of Support (Maria) before issuingrefund-policy.md §7
Proposed decision
Partial refund — $533.33ProposedQueued in Actions — awaiting your approval

Hit a gap in the policy and the agent escalates instead of guessing — and issuing the refund always stays a human step.

Deal-desk — Acme Corp · 22% discount requested
Matched policy rule
Annual plan, discounts ≤ 15% auto-approve at rep leveldeal-desk.md §2
15–25% requires Deal Desk sign-offdeal-desk.md §3
Guardrail
Escalate: 22% is above the 15% auto-limit — route to Deal Desk before quotingdeal-desk.md §3
Proposed decision
Escalate to Deal DeskProposedSlack alert drafted — awaiting your approval

Same engine as refunds — but this workflow has its own decisions (approve · escalate · deny) and its own 15% limit, both read from your policy.

Access request — prod database · read-write
Matched policy rule
Read-only access to non-prod → grant with manager on recordaccess-policy.md §1
Production read-write → requires security owner approvalaccess-policy.md §4
Guardrail
Escalate: prod read-write is above self-serve — needs security owner (Dev) sign-offaccess-policy.md §4
Proposed decision
Escalate — open Linear issue for DevProposedIssue drafted — awaiting your approval

The agent drafts the Linear issue and assigns the owner — but granting the access stays a human decision.

The compounding moat

A brain that gets sharper every time.

Most tools are as good on day 100 as day one. RampBrain isn't. Every time your team resolves a case the agent escalated, that resolution becomes part of the brain.

  • 1The agent hits a gap in your policy and escalates instead of guessing.
  • 2A human resolves it and captures how — one note, a few checkboxes.
  • 3RampBrain stores it as a sourced document and recompiles the skill in place.
  • 4The gap is closed. Next time that case comes in, the agent just handles it.

The work gets more consistent, and the brain more valuable — exactly the asset that's hard for a competitor to copy, because it's built from your resolutions.

the brain
Decision Resolution Recompile Gap closed
It doesn't stop at a decision

The agent does the safe work itself.

Once you approve, RampBrain takes the coordinating action — so the decision turns into motion, not another task on someone's list. The irreversible steps stay yours.

Posts to Slack

Fans out a grounded heads-up to the right channel — the matched rule, the proposed call, and the source — so the team sees it where they already work.

Safe · still gated

Opens a Linear / GitHub issue

Drafts the issue with the owner, context and citations attached — the escalation becomes a tracked ticket the moment you say go.

Safe · still gated

Calls your webhook

Fires a clean payload to your URL — n8n, Zapier, your own automation. We store no third-party secrets; the action posts only where you point it.

Safe · still gated

Money moves and production changes never auto-run. They stay a cited human checklist with a hard gate in code — an action that needs approval and hasn't got it is blocked before it can touch anything.

What Makes Us Different

Not a chatbot. A system that does the work.

Everyone can build RAG over documents. Our moat is two layers deeper — and it compounds.

Executable Skills

We don't just answer. We compile your knowledge into skills an agent runs.

Any policy becomes a grounded SKILL.md — decisions, steps and guardrails, each traced to a source — that an AI agent loads over MCP and actually executes. Q&A is table stakes; doing the work is the product.

Grounded · Gated · Self-updating

It acts only on your rules, never without sign-off — and it learns.

It works from documented rules only and escalates on a gap. With your approval it runs the safe coordination; money and production stay a human checklist. And every resolution you capture feeds back into the brain — so it's sharper next week than it is today.

app.rampbrain.com
The RampBrain app deciding refunds: a compiled refund-policy skill with approve / partial / deny rules, each grounded in the team's policy and flagged where a human must approve
A real compiled skill in the app — a refund policy turned into grounded rules (approve · partial · deny), every step cited to its source, with the money-moving steps gated for human approval.
Security & trust

An agent you can actually trust to act.

Letting AI do real work only works if it can't go off-script. Your knowledge stays yours, and nothing the agent proposes runs without you.

🔒

Encrypted & isolated

Everything moves over encrypted (HTTPS) connections, and any connector tokens are encrypted at rest. Each workspace is fully isolated — your knowledge is never mixed with another company's.

Human-gated, nothing auto-runs

Every action waits for one-click approval. Dry-run actions simulate safely; outward actions (Slack, Linear, your webhook) only fire on your go — and money or production steps are hard-blocked in code without it.

🗑️

Yours, never sold

We never sell your data or share it for advertising. You choose what gets ingested and can delete any document, source, or your entire workspace at any time. Delete it, and it's gone.

Use Cases

Real work, run from your own playbook.

The operational calls your team makes from memory every day. RampBrain runs them from your documented rules — sourced, gap-aware, and waiting on your approval.

A customer wants a $640 refund — is it within policy?

The payment service is down at 2am — what fixed it last time?

Sales wants a 22% discount — who signs off, and is it allowed?

Someone needs prod database access — grant it or escalate?

This case hits an edge the policy doesn't cover — escalate, don't guess.

A new engineer starts Monday — draft the day-one plan, then wait for me.

Pricing

Start free. Upgrade when agents do more.

Free to compile your first skills. Go Pro for unlimited workflows, every connector, and agents that run the work end-to-end.

Free
$0 / forever
For trying it on a single repo or team.
  • 1 connected source
  • 3 skill runs / month
  • Knowledge map & Ask
  • Tribal-knowledge capture
  • Community support
Start Free
Enterprise
Custom
For larger orgs with security & scale needs.
  • Dedicated AI quota / your own keys
  • SSO & role-based access
  • Self-host option
  • Security review & SLA
  • Dedicated support
Contact us
No credit card required to start. Cancel anytime. Prices in USD.
* Unlimited under fair use — generous limits sized for normal team usage, not automated bulk runs.
The biggest blocker to AI automation is no longer the models. Now the blocker is the domain knowledge. Every company in the world is going to need a company brain.
Tom Blomfield — Y Combinator, Request for Startups
Jay Gautam, founder of RampBrain

I'm Jay, a final-year CS student. I kept seeing the same thing — the knowledge that actually runs a team lives in a few people's heads, so every refund, every incident, every approval waits on whoever remembers the rule. RampBrain is my attempt to fix that: turn what your team already knows into skills an AI agent can actually do the work with — safely, always under your sign-off, and getting sharper every time. I'm building it in the open, and I read every piece of feedback personally.

Jay GautamFounder, RampBrain
FAQ

Questions, answered.

How is this different from ChatGPT or a doc search?

ChatGPT and search surface a document — they don't do the work. RampBrain compiles your knowledge into grounded, governed skills an agent runs: it acts only on documented rules, refuses to guess (it escalates) when a request hits a gap, traces every step to a source, and stops for human approval before anything executes.

What kinds of work can it run?

Any operational decision your team makes from a policy — refunds, incident response, discount and deal-desk approvals, access requests, spend sign-off, vendor review, onboarding. You don't get a fixed list of features: you point RampBrain at a policy doc and it compiles a workflow with that policy's own decision options and limits.

Is it safe to let an AI handle refunds and incidents?

Yes — by design. The agent only acts on rules documented in your sources, escalates instead of guessing on a gap, and every action waits for one-click approval. Money decisions and production-mutating steps stay a cited human checklist, never auto-run. More in Security & trust.

Does it get better over time?

Yes. When a case is resolved, you capture how — RampBrain stores that as a sourced document, recompiles the skill in place, and closes the gap that caused the escalation. The brain gets sharper the more your team uses it.

What can I connect?

GitHub repositories, documents (PDF, Markdown, code files), Slack exports, Jira, Confluence, Notion, Linear — plus 5-minute knowledge interviews. And outward: the agent can post to Slack, open a Linear or GitHub issue, or call your own webhook.

Who is it for?

Teams that run the same operational decisions from memory — support handling refunds, on-call responding to incidents, ops approving discounts and access — and want them done consistently, sourced, and safely.

Put your company's
playbook to work.

It's live and free — connect a repo, compile your first skill, and watch an agent propose its first action. You approve.

Or get product updates:

Occasional product updates — no spam, unsubscribe anytime.