Personal engineering build
Several small devices, each with a microphone and speaker, all streaming audio to one server. The server works out who said what, matches it against known commands even when the words are not exact, and triggers the right response on the right device.
Devices hold an open websocket to the server and stream audio continuously, so the response starts as soon as someone stops speaking rather than after a file finishes uploading.
Voice activity detection runs on the device itself, so silence is never sent over the network. This alone removed most of the load from the server.
Recognised text is matched against known commands with fuzzy string matching rather than exact comparison, so near-misses and accents still work instead of failing silently.
Each device has an identity and its own mapping of what to play, so the server can answer on exactly the one that was spoken to.
A four-branch pharma distributor ran inventory in one system, billing in another and accounts in spreadsheets. Batch and expiry tracking existed only in a storekeeper's memory. We consolidated onto a single ERP with batch-level traceability and automated the reconciliation that three people were doing by hand.
Selling on its own website plus four marketplaces, the brand was overselling roughly 4% of orders and eating cancellation penalties and rating damage. We built a single inventory truth with near-real-time propagation to every channel, plus automated settlement reconciliation.
Production planning ran on a whiteboard, job cards were paper, and quality data was typed into Excel twice. We digitised the shop floor flow and automated the reporting layer above it, without replacing the ERP the company had already invested in.
Most of these started with a 30-minute call and an honest diagnosis.