Engagement · Cadence
Hardened mobile data collection, high-speed ingestion, and an Auto-ML baseline. Production-grade infrastructure for behavioral signal at scale.
The problem
The existing digital phenotyping stack was built in academic labs to answer a single paper. It buckles when you push past prototype. We aren't iterating on it. We're rebuilding underneath it.
Beiwe. Aware. RAPIDS. Open-source tools that grew up in academic labs to answer a paper, not run at scale. Brittle SDKs, lossy collection, no clean path from raw signal to a usable feature.
Hardened mobile capture across iOS and Android. Lossless ingestion at scale. Features that modellers can actually work with. A provenance chain everyone in the loop will trust.
A native-mobile, AI-native infrastructure for digital phenotyping. Built to a 13-week clock, with each milestone a working surface the team can run against.
The pipeline
Native capture, lossless ingestion, modellable features, AutoML baseline. One pipeline. One source of truth. One handoff.
Milestone deep dive
Swift and Kotlin SDKs written the idiomatic way, not auto-generated. Battery-aware sampling, lossless local buffering, structured retry. The collection layer survives airline mode, low-power mode, and OS upgrades.
Backend pipeline tuned for thousands of signals per second per participant. Schema-enforced at ingest. Cleansing rules versioned with the spec. Throughput tested at full-cohort scale before sign-off.
Feature engineering pipeline that turns raw signal into the columns the modellers actually want. Flutter demo app for live participant inspection. Operations dashboard an operator can run unattended.
Auto-ML baseline that flags initial behavioral patterns in the sample. End-to-end test of the full pipeline. Documentation the team uses to operate the platform without us in the room.
Timeline
Foundations
Digital phenotyping is the next instrument for measuring how people actually behave. The signal is rich, but the existing tools were never built for the load. The platform dies on infrastructure, not on the science.
Sensor coverage on consumer phones is finally high enough to drive real pattern discovery. The bottleneck is no longer the device — it's the pipeline between the device and the modeller.