Build

Line-of-business applications, integrations between systems that were never meant to meet, and the screens your people actually touch.

At the automation end this means agent systems: software that works through a queue of operational tasks, keeps a record of what it did and why, and stops to ask before it spends money, sends anything, or does something it can't undo. The stop-and-ask is the design, not a limitation.

Typical work: Intake that reads the PDF; the Friday retyping job, deleted; a back-office queue that works itself down and asks before it acts.

Most businesses don't need a general chatbot. They need one task handled reliably: reading a type of document, sorting requests, drafting replies, making a routine call. That is often best done by a small language model (SLM), one compact enough to run on a single workstation or on the device itself, fine-tuned for exactly that job. Where a large model (LLM) is the right tool, you'll be told that too.

The work covers the whole path: choosing the model, preparing the training data, fine-tuning (including LoRA, which adapts a model cheaply without retraining all of it), and compressing it so it runs on ordinary hardware without losing what it learned. Behind that sits research into steering: adjusting how a model behaves directly, instead of pleading with a prompt.

Every result is measured on cases the model never saw, with the uncertainty stated, and the models we publish say plainly where they fall short. You get a number you can check, not a demo. When the data is sensitive, the model can run on hardware you control, so the data doesn't have to leave the building.

Typical work: A private assistant for one workflow, on your own machine; a model fine-tuned on your documents; an honest test of a model or vendor you're considering.

Voice that drives software directly, rather than a transcript somebody has to read. The hard part is time: working out what was meant while the person is still talking, on the device, inside a latency budget measured in milliseconds. That budget gets measured, not assumed.

Typical work: Hands-busy control for a workflow; voice on a device with no usable screen; speech handled on your own hardware instead of a cloud service.

Most businesses don't have a data problem. They have six systems and no way to ask them a question at the same time. Sometimes the answer is one dashboard that replaces a monthly spreadsheet ritual.

At the other end are pipelines that refresh daily and merge only what changed, datasets assembled and documented from public and licensed sources, and quantitative and financial analytics. The engineering that matters is correctness: catching a record that was edited after the fact before it gets counted as though it was always true.

Typical work: A dashboard that replaces a spreadsheet ritual; a dataset built, cleaned and documented; an audit of the number everyone quotes.