Kilow platform

The problem

The backend was in poor technical shape and the business kept changing what it needed. The job was to keep up with those requirements while improving the system underneath them, before AI could really help with development.

A meal from a photo

How a patient logs a meal or asks about their plan on WhatsApp, without opening the app.

1. 01 · WhatsApp

   A patient sends a photo

   Patients talk to the Kilow assistant on WhatsApp: a photo of their plate, or a question about their plan.

   No app needed
2. 02 · Intent detection

   The assistant works out what they want

   Vector-backed intent detection tells a meal to log apart from a question to answer.

   Qdrant · OpenAI
3. 03 · Memory

   It remembers the conversation

   Each conversation keeps its own memory, so a follow-up question makes sense without starting over.

   Per-conversation memory
4. 04 · Patient record

   The meal is logged

   The meal is logged from the photo against the patient’s plan, alongside their trackers and goals in the app.

   Meal logging from a photo

One job system for everything

How every scheduled task, from a medication reminder to a subscription renewal, runs through one system the team can see and manage.

1. 01 · Scheduling

   Every future task is a job

   Reminders, renewals and follow-ups are delayed jobs stored in the database, so the team can see, move or cancel them.

   Delayed jobs in PostgreSQL
2. 02 · Queues

   Each kind of work has its queue

   Payments, reminders, notifications and analytics each run in their own queue.

   BullMQ · Redis
3. 03 · Payments

   Subscriptions follow one set of rules

   Trials, renewals, refunds and extensions live in one place, and each payment runs through a pluggable strategy for its method.

   Card · Apple Pay · buy now, pay later
4. 04 · Accounting

   Receipts reach the books

   Every payment generates a receipt that syncs to the accounting system.

   Receipts synced automatically

Key decisions

1. 01

   Support became a real-time CRM: tickets with comments, mentions and attachments, pushed live to every agent, with AI-written summaries of each request.

Outcome

Took a backend that was a mess technically and made it count: a system the business could keep building on, until the company stopped operating in March 2026.