The Data Factory Operating System
The premise
David AI's stated goal is to chase capabilities in the agentic-voice space that don't exist yet. This includes identifying cutting-edge classes, designing the shape of the data, proving it internally, and then scaling towards a product offering.
I'm offering an operating system that decides which hypotheses to fund, converting them into buildable specs to test against, and then industrializing once scalable value is identified. The capability I'm focusing on is agentic tool-grounded voice, with the models and dashboards I'm providing running through a worked example end-to-end. That said, every stage is capability-agnostic and can operate with any parameters without changing the underlying structure.
The question
How do you make pilot design and scaling reproducible, instead of re-deriving the economics from scratch for every new dataset?
The worked example
An office voice agent that works through problems via tool calls: spoken requests paired with a time-aligned trace of the actions they triggered.
The output
A model for what an hour of that data costs, how many people it takes, and in-flight quality monitoring.
- $1,548
- per delivered audio hourall-in, at 1,000 hours
- $103.22
- per interaction≈4 min, 6 tool calls avg
- $17.20
- per tool-call label90 labels per delivered hour
- $155k
- recommended pilot12 capability classes × 100 hours
The operating loop
Every new data product moves through five stages, with graduate-or-kill governance spanning the middle: intake turns an open-ended researcher hypothesis into a scored brief; data-shape design specs the collection rig, label schema, and QC gates together, since the label you can afford constrains the capability you can sell; pilot/de-risk ranks what the pipeline assumes and spends on the cheapest experiment that could kill it first, before committing to a kill box; industrialize and operate then take a validated shape to scale and keep it instrumented.
The two stages with real numbers behind them, pilot/de-risk's cost model and industrialize's staffing plan, are worked out in full under RTF Model and Capacity Model above. The rest of the loop is real but stays at this altitude here.