Nirmano — turning a factory's ERP into answers

A mid-market manufacturer runs on its ERP, but almost never gets answers from it. The data is all there — orders, invoices, production runs, quality checks, receivables — and yet "are we collecting slower than last month?" or "which jobs are quietly stuck?" still takes a spreadsheet and an afternoon, if anyone has the time at all. The insight exists; it's just locked up.

Nirmano is the product I'm building to set it free. It connects to a manufacturer's ERP and turns it into three things they didn't have: a daily briefing that lands in the morning, a copilot you can ask questions in plain English, and quiet alerts when something starts to drift — receivables aging, a job stalling, inventory slipping below where it should be. It's running in production today for a live mid-market manufacturer.

The interesting part wasn't building dashboards. It was earning the right to put AI in front of someone's business at all.

Answers you can act on

In a finance or production meeting, a number that's almost right is worse than no number — people make decisions on it. So the hardest requirement wasn't intelligence, it was trust: the system can never quietly invent a figure.

The way I solved that was to keep the AI away from the parts where being wrong matters. Whether receivables are overdue, whether a work order has gone stale, whether stock is below its floor — those are settled facts, decided by plain, testable rules, never by a model's judgement. The AI's job is the last mile: taking numbers that are already correct and turning them into a briefing a plant manager will actually read. Every figure it writes has to trace back to real data, and an automated check rejects anything that doesn't before it can ever ship. The result is a system people came to rely on, because it has never put a made-up number in front of them.

AI that can't run away with the budget

The other fear with autonomous software is the runaway — an agent that loops, burns money, or fires off an email no one approved. Nirmano's agents work inside hard limits. Each one has a ceiling on how long it can run, how many steps it can take, and how much it's allowed to spend, and it stops itself the moment it hits one. Nothing goes out the door on its own, either — every email or ticket it produces is a draft for a person to approve and send. The payoff is boring in the best way: predictable cost and no nasty surprises, which is exactly what makes it safe to leave running every day.

Making a messy ERP usable

Real ERP data is never tidy. One customer ran their system across two separate databases where the same code meant two different things; the signal that makes manufacturing analytics worth anything was buried in custom fields the ERP treats as afterthoughts; and even "revenue" hides a tangle of tax and currency. A lot of the work was the unglamorous translation that turns all of that into one clean, consistent picture a manager — and the AI — can reason about. None of it is visible in the final product, which is the point: the mess stays on my side of the line.

Where it stands

Nirmano is live: the daily sync runs, the briefings go out, the copilot answers real questions against real data, and the alerts surface problems before they turn into fires. It's proven with one manufacturer on one ERP; the next step — and the real test of the design — is the second.

What I'd point to here isn't a clever model. It's that this is AI doing real work in a real business, built the way you'd build anything people depend on: trustworthy where it counts, kept on a short leash, and honest about what it can and can't yet do.