Build
RR Mart
Retail, small-town and rural grocery
One command runs the chain from data. Roughly 50 agents read ERP exports and store WhatsApp chats on a local machine and return a dashboard, an Excel file and a short summary, with every number tagged to its source.
Nothing was measured after delivery. No one set up the measurement, so the honest answer is that the outcome is unknown.
The problem
A growing chain of small stores, run from one warehouse, kept its operating reality in two places: an ERP system, and WhatsApp groups. Both were full of answers and neither would give them up. The owner could not easily see which stores were overstocked, which products were close to expiry, or which stores were falling behind target.
This is the ordinary shape of the problem in a small business. There is no data team to ask, and no appetite for a system that needs one.
What I did
I designed and built a single command system for running the chain from data. The owner picks one command. The system reads the ERP exports and the store WhatsApp chats on a local machine, and returns three things: a dashboard, an Excel file, and a short summary written in plain language.
Underneath, roughly 50 agents handle the specific jobs: dead-stock detection, near-expiry detection, transfers between stores, stock-out detection, daily briefings, store ranking, and alerts.
Two decisions matter more than the agent count.
It runs locally. The chain’s sales and chat history never leaves the owner’s machine, which is both a privacy answer and the reason it was allowed to touch WhatsApp data at all.
Every number is tagged to its source. If the briefing says a store is 12% behind, the owner can see which export and which rows that came from. In a business where the ERP and the chat sometimes disagree, an untraceable number is worse than no number.
It has been run repeatedly since June 2026. I also set up a LinkedIn brand system for the chain.
What came of it
No outcome was formally measured. The system runs and is used, and I have not measured stock-outs avoided, write-offs reduced, or hours saved, so I am not going to imply a figure.
The systems behind it
What I wrote about it
- Cheaper tokens, broken budgets: why AI agents need spending caps Google now lets a hard spending cap pause its work agent, which tells us agents should be run as cost centres, with an owner, a limit and a named person who can press resume.
- Where enterprise AI stalls: between the pilot and daily use Anthropic's $100 million academy suggests the hard step in enterprise AI is getting from a pilot that works to a system people use every day, and that depends on people more than tools.
Other build work
- A scientific publishing services firmTwelve ranked automation opportunities, one shared architecture, and working prototypes: a reference and retraction checker, a books project-management system, and a rules-first manuscript-to-XML pipeline.
- An engineering design firm for building servicesFour hours per floor go into cleaning an architect's drawing by hand before engineering can start. I built and measured a prototype that does the tractable parts of that, and reported the part that is not solved.
- InticoreA quotation-rules knowledge pack for a custom GPT that reads a casting drawing, tabulates every dimension and tolerance, checks feasibility and writes the queries back to the customer.
Have a problem worth measuring?
I take a small number of advisory and build engagements each year.