The shape of the thing
YourKith is an intelligence cycle.
Not a database with reminders bolted on. Turning raw signal into a decision has a shape, and it has had the same shape for a long time: you decide what you need, you collect it, you make it usable, you turn it into judgement, you get it to whoever acts on it, and then you let what happened next change how you do all of it. Six stages, one loop. Every part of this platform sits at a named stage, including the parts that are still thin.
01
Direction
Somebody has to say what matters
You define who you're trying to reach, and you set one number, the "working with" window, that decides what counts as close. Every band below it is a fixed multiple, never a second dial that can drift out of agreement with the first.
Which modules run, which sources connect, and what the platform is allowed to spend are deliberate, admin-set decisions, and the governing rules are readable rather than folklore.
The honest read: today this stage shapes how signal is scored and ranked once it arrives. It does not yet decide what gets collected. Connectors sweep a time window and the profile sorts out what landed. That ordering is the wrong way round, and it is why relevance has to be re-decided by a gate at every downstream seam.
02
Collection
Signal from where it already lives
Connectors drain on a schedule and write with provenance, so every row knows where it came from. You do no data entry; the corpus is the exhaust of work you were doing anyway.
The honest read: collection is the youngest and least even part of the cycle. Reasoning can only ever be as good as the roster it was handed, and a meeting captured before we persisted its attendee list is a meeting the resolver cannot reason about.
Live today: Gmail · Google Calendar · Slack · Zoom · Granola · Fathom · Mercury · QuickBooks
On request: Outlook · Outlook Calendar · HubSpot
03
Processing
Raw signal becomes a resolved person
This is where most relationship tools quietly fail. Identity lives in one durable identifier index, and employment history is append-only, so it is never overwritten. Records fold together only on a shared verified, non-role identifier, because a wrong merge is the one mistake a later pass cannot detect and undo. A matching name is never enough.
And nothing reaches a shared surface ungated: every source runs through a declared, tested relevance gate before it can put a single row in front of you.
Durable identifier index · append-only employment · merges adjudicated by a human
04
Analysis
Resolved entities become judgement
One classifier turns recency and tie-existence into a single word, working, cooling, dormant, or network, and the graph, the directories, and the health scoring all read that one value. There is no second definition to disagree with it.
On top of it sit company and contact intelligence, ratings, briefs and competitive battlecards. Every enriched fact must carry a source; an unsourced claim is held rather than quietly committed.
One tie-state classifier · provenance gate on every enriched fact · explainable scoring
05
Dissemination
Put it where the decision gets made
A daily briefing, one insights surface, the directories, the relationship graph, and a single triage queue where everything that needs you lands. One inbox, not one per source.
And Wita: chat over your own corpus. When an answer is grounded in your records it carries numbered citations that deep-link to the record each one rests on, and a citation that cannot be resolved to a real page renders flagged rather than being given a fabricated link. When there is nothing to cite, it says it does not know instead of guessing.
Briefing · Insights · Relationship graph · Triage · Wita, with resolvable citations
06
Feedback
A loop, or it's only a pipeline
Every triage decision is recorded. A source you keep dismissing as genuinely irrelevant stops filling your queue, but only once it has proven noisy on a real sample, never off a handful. Dismissing something because you had already handled it never counts against the source; it was relevant, you just dealt with it.
Reject a proposed merge and that pair is never proposed again, and the thresholds tighten upward only, never loose enough to let a bad merge through.
The honest read: feedback is real and wired, but it is the newest stage. Today it learns to suppress and to tighten. Learning to promote is ahead of us, not behind.
Why say it this way? Because it is the difference between a tool that stores what you typed and one that closes a loop. It also makes the weak links findable instead of flattering, on this page, in this order, you can see which stage we would defend and which one we are still building.