One dashboard that runs a flower supply chain.

The architecture has one organising idea: the database is the truth, the AI is the narrator. Every night a job walks the entire operation and writes a compressed, human-readable memory — eleven categories covering clients, cities, flowers, cashflow, festivals, labour, wastage, farmers, pricing and the company's North Stars. The next morning the AI reads that memory instead of re-querying everything, which keeps the briefings fast, cheap, and consistent across the team.

It knows the craft: garlanding is the bottleneck, so it starts every garland SKU at the top of the shift; loose flowers get weighed before sealing; whole items like lotus and betel leaf are just counted into a bag. As packets get sealed, the floor supervisor taps a block and logs progress — and the AI re-plans around what's actually left, shifting people to wherever the day is running behind.

The dashboard steps through the page's own numbers one at a time — each with its target, a ✓ or ✗, and a one-tap button to raise a task the moment something's off. The agenda writes itself from the data, the discussion stays anchored to real figures, and the action items are already on the right person's dashboard before the meeting ends. Combined with the meeting-to-task pipeline below, nothing said in a room gets lost.

The model is intentionally interpretable rather than a black box — a buyer has to trust it at dawn. It starts from a seven-year price history for each flower variety in each city, builds a seasonal baseline (jasmine in May behaves nothing like jasmine in December), layers a festival multiplier from the Hindu calendar (demand spikes around Varamahalakshmi, Navaratri, Diwali are learned, not guessed), and then nudges the whole thing with the last few days of live market prices the team logs. Each prediction ships with a confidence band, so a buyer knows when to trust it and when to haggle.

Every meeting is recorded through Fireflies, which transcribes and extracts action items. A job syncs those transcripts into the database every thirty minutes, where each action item is matched to a person and written out as a task with an assignee, a due date and a link back to the meeting it came from. It surfaces on that person's front page, feeds into their personal daily plan, and the AI CEO can see it when reasoning about whether someone is keeping up. Talk in, accountability out.