The weekly desk
About five account cards a week by default, fewer when fewer pass the checks. Each card says who, why now, and which buyer role to approach, with the evidence attached.
Zeki product · AI sales intelligence
DealBell
DealBell finds companies with two dated, sourced reasons to buy now, and keeps the evidence for every claim. It is our own product: an AI sales intelligence engine that does the research a good SDR does, then hands the account to a person to decide.
Type
Our own product
Shape
Signals, evidence, gates
Output
A weekly desk of accounts
Sends email
Never on its own
Status
Pre-launch, by request
The problem
What it replaced.
Outbound tools are good at volume and bad at reasons. A rep gets a list, a template and a guess about timing, and the buyer can tell. The research that would make the message worth reading, who is hiring for what, what changed last month, who actually signs, takes an hour per account, so it rarely happens.
What we built
The pieces, and why each one exists.
An evidence trail for every claim
Each statement links to a saved copy of its source and date, and is tagged as stated or inferred. A claim the sources do not support is dropped rather than softened.
Five gates before an account ships
Fit, why now, can they buy it from outside, is the door open, and is there a named buyer. A single signal is held on a watch list until a second, independent, dated signal arrives.
Exclusivity and a person in charge
Each account goes to a limited number of customers, with a receipt. DealBell never sends anything itself: the customer's team decides and sends from its own tools.
The hard part
Where the time actually went.
The hard part is honesty at scale. A model asked to explain why a company might buy will always find a reason, so the system is built the other way round: it starts from dated public evidence, refuses to ship an account on one signal, and drops any sentence it cannot trace back to a source. Its crawler reads public job posts through the job boards' own public APIs, once a day, rate limited, strips emails and phone numbers, respects robots.txt and honours opt-outs.
What we took from it
Carried into later work
- Evidence first, model second: the model writes from sources, never instead of them.
- Two independent signals beat one strong one.
- A human decides every send.
The stack
What it runs on.
Sources
Public job board APIsCompany newsFilings and announcements
Checks
Dated signal extractionTwo-signal ruleFive gates
Product
TypeScriptNext.jsPostgres
Data rules
No LinkedIn dataNo personal contact data storedrobots.txt respected
Worth knowing
A product we use ourselves.
DealBell is built and run by Zeki, and it is the same kind of system we build for clients: AI that does the research and the drafting, with evidence attached and a person making the call. It is pre-launch and available by request at dealbell.ai.
See it live: dealbell.ai.
Building something like this? See how we work as an AI implementation partner.
Next step
Bring us the problem, not the spec.
Thirty minutes. If your problem rhymes with this one we will say what we would do differently the second time.