A conversational data agent that turns factory sensor data into answers.

Customer background

Alchemi instruments factory production lines with IoT sensors that track equipment performance, material flows, and energy use. That telemetry powers automated reporting on OEE, mass balance, and energy efficiency for FMCG and industrial clients — the lines producing familiar goods like tubs of yoghurt, boxes of beef patties, and blocks of cheese.

The platform already captured rich, real-time data. The open question was whether Alchemi's own customers could actually use it — without raising a report request or waiting on a stretched data team.

The problem

Useful data existed, but the insight inside it was effectively invisible. Answers that mattered took too many clicks and filters, and the people who needed them rarely had the skills or access to get there.

Escalation was reactive and manual. When something looked wrong on a line, it often came down to a manager taking a cellphone photo of a screen and passing it on. Dense daily reports only made it harder — they buried what mattered, so recurring patterns went unnoticed and few people could turn the raw data into root-cause analysis.

Low efficiency scores drove real waste. Problems on the production line often went undiagnosed for days, with no systemic way to track issues against KPIs — and by the time they surfaced in a report, thousands of dollars of product would already have been wasted.

The solution

We built a customer-facing, LLM-powered self-service data agent — one their users could simply talk to. A model pointed straight at a raw database won't get there on its own, so most of the engineering went into the layers around it that make it dependable:

• Plain-language queries return dynamic charts and contextual recommendations, not raw query dumps.
• A materialised data layer over Alchemi's existing sources, reshaped so the agent reasons over clean, unambiguous data.
• Multi-tenant safety by design: every query routes through authenticated MCP endpoints, so no customer can ever see another's data, and the agent never writes raw SQL against the database.
• Near real-time alerts routed by frequency and severity to Slack, Teams, or WhatsApp.
• Continuous evaluation: an eval framework and user-feedback loop grade the agent against known-correct answers and keep it improving in production.

Crucially, the agent earns trust the way a good analyst does — by being measured. A continuous evaluation suite scores its answers against results we'd verified by hand, and the gaps it surfaces tell us where to tune the model and where to fix the data underneath it.

Alchemi Data AgentLive

The impact

The right data now reaches the right user, at the right time, through the right channel — with full ability to export, share, and query further. Problems that once surfaced in a weekly report are caught and actioned in minutes.

The agent functions as a skilled analyst: valuable to Alchemi internally, and commercially viable as a standalone product they can sell to their own customers.