No case studies yet. Six systems in production.
We would rather show you what the software does than tell you what a client thought of us. Everything below is running now, described in enough detail that an engineer could check it.
Client names withheld · the shape is real
A lending platform, replacing a Salesforce org
The org could not model the process. Positions, stipulations, parallel lender submissions and commission splits each needed a place the object model did not have, so the workarounds had become the system.
Assignment weighs live workload, tier and six months of conversion rather than distributing in rotation. Duplicates awaiting review are excluded from capacity, so nobody is penalised for a vendor's bad data.
Score rows are stored and the total derived from them. There is no code path that writes a bare score, which means any decision can be explained after the fact.
Consent, suppression and quiet hours are defaultChecked before every attempt, in code rather than in policy, because TCPA damages are per message.
Rep, manager and ISO shares depend on which lender funded and at what buy rate, so the arithmetic lives on the record rather than in a month-end spreadsheet.
A partner lead-ingest API
Partners were emailing submissions and discovering rejections days later. The API answers in the same request, so a partner knows immediately whether a lead was accepted and why it was not.
The receiving end validates, deduplicates and answers before the request closes, which changes the partner relationship more than any dashboard does.
Email misses most duplicates in this market. Matching on business identity and address does not.
A deduplication key on every event, so a retry creates one record rather than two.
Already a customer, revenue missing, asked not to be contacted — reasons a partner can act on rather than a generic failure.
Server-side attribution to five ad platforms
The platforms were optimising towards form fills because that was the last thing they were told about. Funded deals were arriving weeks later and never reaching the systems spending the money.
Qualified, application and funded derived from what actually happened rather than assembled by hand, so the same definition produces the same number every time.
Events leave our infrastructure rather than the visitor's browser, surviving blockers, consent refusal and a thirty-day sales cycle.
A high-water mark per platform, so a re-run never sends the same funded deal twice — which is very hard to unwind once it has happened.
In one view, so the comparison is unavoidable rather than optional.
An SMS assistant that stands down
The first three messages of a conversation were consuming rep time on every enquiry. The assistant handles them and hands back the moment a person should have it.
Human reply, escalation request or message ceiling — defaultChecked in code before every outbound message, because a model can be talked out of a prompt.
Each answer maps to a field on the lead record. If there is nowhere to write it, the assistant does not ask for it.
There is no field in the schema to write one to, so the constraint is structural rather than persuasive.
Tokens and spend recorded per thread. An agent whose cost nobody can see is an agent nobody controls.
What we will not put on this page.
Everything above belongs to the client who paid for it.
The repository is theirs
Created in their organisation from the first commit — not transferred at the end, which is the arrangement that goes wrong.
No software commission
No commission, referral fee or partner margin on anything we recommended while building any of it.
Export rights, always
Their data leaves in a usable form whenever they ask, without a project to make it possible.