Someone here has already solved it.
Keegi is for organisations where nobody can hold the whole picture in their head any more. Ask who can help, in your own words, and you get three colleagues who have actually done the thing, each with the evidence for why.
Twenty minutes. No demo, no pitch. I want to know whether this problem is real where you work.
Keegi never tells a manager how much someone works. It tells a colleague who to ask. How that is enforced.
Querying Northwind, a synthetic company we use for demos.
Marta Kowalski · Programme Lead · Berlin
Ran the Betriebsrat consultation for the Berlin office’s Workday rollout in 2024, including the two sessions that changed the logging scope.
Agreed with the council that endpoint telemetry is out of scope for phase 1; revised the data map accordingly before the second session.
DOC-1193Drive · Mar 2024
Jonas Herrmann · IT Director · Munich
Negotiated the works agreement covering endpoint monitoring ahead of the M365 migration, a year before the Berlin rollout.
Aoife Byrne · People Ops Lead · Dublin
Wrote the DPIA the Berlin council reviewed, covering the same rollout from the data-protection side.
The problem
Right now, somebody asks the channel.
They get two replies. Both are helpful and neither is from the person who actually did it, because that person moved teams a year ago and stopped reading the channel. The thread scrolls away. Someone asks the same question next quarter and the same two people answer it.
The person you needed was in the building the whole time. Nothing you own knows how to point at them.
How it works
Nobody fills in a skills form.
Connect your sources.
Read-only. Every source starts excluded, and an admin switches each one on.
Keegi reads what people shipped.
It writes one card per person from pull requests, docs, resolved threads. A claim without a citation gets dropped. Density counts artefacts, not words.
Then ask.
No filters or tags. At most three people, one sentence each, and a drafted opener you edit before you send it.
What it reads
- Pull requests
- Design docs
- Resolved threads
- Sources it may not quote
What it writes
Priya Raman · Senior Engineer, Payments · London
Rebuilt the payments read replicas inside the same upgrade window and verified reconciliation afterwards.
Evidence
You can open every sentence.
Each card is a claim, and the artefact it came from sits underneath it. Open the chip and you get the paragraph, the system it came out of, and the date. A sentence Keegi cannot trace to something a person produced does not get softened, it gets dropped.
Tomas Sepp · Platform Engineer · Tallinn
Ran the Postgres 13 → 15 upgrade of the primary cluster using logical replication, with a 40-minute cutover.
Lag gate was set on the wrong publication. Caught before traffic moved; the fix is in the runbook above.
PM-0355Drive · Jun 2024
Why this person
There is a runbook with their name on it, and a postmortem saying the cutover ran twelve minutes long. The postmortem is the reason this name comes first. Nobody writes “the upgrade overran, and here is what I got wrong” onto their own profile, which is exactly what makes it the best evidence in the corpus that they have done this under real conditions rather than read about it.
The same rule is what produces the refusals. When the corpus does not support an answer, Keegi says nobody and logs the gap rather than returning its closest guess. The console at the top of this page runs into one every third question.
How we check that it is working: what we measure, and what we cannot show you yet.
Use cases
Ask it the way you'd ask a colleague.
The new joiner
who set up our Stripe webhooks?
The engineering lead
who has migrated a Postgres major version in prod?
The chief of staff
who has dealt with a German works council?
None of these are searches.
Comparison
Keegi, next to what you already have.
| Approach | Answers with | Source of truth | Fails when |
|---|---|---|---|
| Enterprise search | A document | Written artefacts | The doc is stale, covers the wrong case, or was never written |
| The HRIS skills matrix | A skill tag | Self-report, filled in at onboarding | A claim goes stale within a quarter, and managers can see it, so it is gamed upward |
| Asking the one person who knows everyone | A name, from memory | One person’s ten years | They go on holiday, and it does not survive their resignation |
| Keegi | A person, with proof | The artefacts they produced | The corpus is thin, which we measure before selling you anything |
The last row has a page behind it: what we measure, and what we cannot show you yet.
Before a call
Twenty minutes to look at how the company is put together.
Start with where the teams sit if you would rather not book yet. Sites, where the work lives, whether the company grew through acquisition, when the informal network stopped being enough. It does not pretend to be a measurement.
Twenty minutes. No demo, no pitch. I want to know whether this problem is real where you work.
What is for sale is on the pricing page.