GPT-Powered Medication Reminder Agent (MVP)

Problem

Andrey's startup is tackling medication non-adherence — a problem that costs healthcare systems billions annually and affects roughly 50% of patients with chronic conditions. The core insight: most reminder apps feel like alarm clocks. Patients snooze them and forget.

The hypothesis: a conversational agent that checks in, responds to context ("I already took it", "I'm feeling sick today"), and adapts its behaviour is meaningfully better than a push notification.

The goal for the MVP: prove the hypothesis with real users before committing to a full mobile app build.

Solution

A Telegram-based agent built on n8n and GPT-4o:

1. Patient onboarding — The bot collects medication names, dosages, and preferred reminder times via a structured conversation. Data is stored in Supabase
2. Scheduled reminders — n8n cron triggers send personalised Telegram messages at the right times. The message includes the medication name, dose, and a simple "confirm / skip / snooze" prompt
3. Conversational responses — GPT-4o handles free-text replies. "I already took it an hour ago" → logged as taken. "I'm nauseous, should I still take it?" → empathetic response + reminder that the bot isn't a doctor, suggesting they contact their physician
4. Missed dose logic — If no response within 30 minutes, a follow-up is sent. After two missed doses in a row, a caregiver or family member receives an alert
5. Adherence dashboard — Weekly summary sent to the patient and, optionally, their caregiver via Notion-generated report link

Result

MVP delivered in 2 weeks Tested with 12 pilot users over a 6-week period 89% daily adherence rate among pilot users (vs. ~50% baseline for traditional apps)

The startup is now in talks with two private clinics for a paid pilot. A mobile app version is in planning.

What made it work

Telegram was the right channel for the MVP — no app install, no onboarding friction, universally familiar. The target demographic (older patients managing chronic conditions) already uses it daily.

The conversational layer was the differentiator. Users in the pilot reported that the agent felt "caring, not nagging." Getting the tone right — warm, clear, never preachy — was the hardest part of the prompt work.

Stack

• n8n (self-hosted) — scheduling, orchestration, alert logic
• GPT-4o — conversational understanding and response generation
• Telegram Bot API — patient and caregiver interface
• Supabase — patient data, medication schedules, adherence logs
• Notion API — weekly adherence report generation