Product
pharmentic.com
Pharmacy POS, ERP, CRM and WhatsApp in one system, run by AI agents with a human approval gate
- Sector
- Pharmacy retail and healthcare supply
- Founder’s role
- Architecture, agent orchestration and full-stack build
Led by our founder, Senesh Fitzroy.
Context
A pharmacy makes hundreds of small decisions a week: what is running low, which supplier is cheapest, whether a delivery matches the order.
Each takes a few minutes. They get skipped when the shop is busy, which is when mistakes cost most.
The problem
- Purchasing is judgement under time pressure.
- Expiry depends on someone remembering to check the shelves.
- A purchase order is real money, so an AI that is usually right is not good enough on its own.
What was built
One system for the counter and the back office
Point of sale, ERP, CRM and WhatsApp messaging share one data model.
Eleven narrow agents, not one general one
Each agent has one job, such as demand forecasting, supplier comparison, purchase order drafting, goods-received checks or expiry management, and a small set of tools for it.
The orchestration loop is written by hand, so the team can read, log and step through every decision.
Forecasting that admits uncertainty
Demand forecasting uses XGBoost over sales history, not a language model.
When the forecast is uncertain, the purchasing agent holds more safety stock instead of trusting a single number.
Security at the database row
Row-level security in PostgreSQL is the authorisation boundary. Agents act with the permissions of the person they act for, so a manipulated prompt meets a database policy.
Architecture
Stack
- Next.js
- FastAPI
- Python
- Supabase
- PostgreSQL
- XGBoost
- GPT-4o
Outcomes
- A working system with eleven AI agents behind a human approval gate.
Recognition
National AI Awards2026
1st Runner-UpNational AI Buildathon 2026Awarded to TEAM POWER, the founder’s hackathon team
What we would do differently
We built the evaluation set after the agents. Several prompt changes felt better and measured worse. On agent projects now, the evaluation set comes first.
Planning something similar?
Tell us what you are working on. We start by understanding the problem, the people, and the systems involved.