Hi, I’m Dmytro: putting AI agents to work on live ERP data

I’m Head of R&D at RigER, a Calgary company building operations management software for oilfield services. Right now most of my attention goes to a new layer on top of our ERP: clients talk to it by voice or text, and agents do whatever the request actually requires inside the system.

That turns out to be hard. An ERP has hundreds of document types, hundreds of master data catalogs and hundreds of reports, and few systems are ready to expose all of that to an agent. Cost becomes a real problem when the data volume is huge, and even working out what the user actually wants takes effort. On top of that, every client configures the system differently, so in practice we end up building dynamic MCP servers whose tools depend on how each installation is set up.

The other half is how we build. About 80% of our development cycle runs through coding agents, with Claude Code orchestrating and Codex implementing. Features are verified by Gherkin scenarios, and any new test step has to get past a separate adversarial agent whose only job is to argue it’s unnecessary. The coding agent never decides what gets tested.

On the side I build AI Video Translator, a Windows and macOS app that transcribes, translates and dubs video. The core idea was simplicity, privacy and autonomy: it can run fully on-device, or combine different local and cloud providers for each step. It also exposes an MCP server, so an agent can translate a video as one step in a larger workflow.

If you’re dealing with large tool surfaces, token cost on big datasets, or agents that misread user intent, I’d love to compare notes. Blunt feedback on the translator is welcome too.