A note from Eduba, for BobiHealth
On-device models. Medically reviewed guidance. A privacy rail pitched at the DoD while three co-founders keep a consumer app alive in three countries. That is a lot of surface area for a pre-seed team. This page is a short read on where the budget should actually land.
What we read before writing this
The thesis
Press coverage in 2024 named three wedges out loud: a consumer subscription, licensing the privacy architecture, and a DoD data rail. The math of a pre-seed team says one of those carries the runway. The other two become ballast until the first one is closed.
That choice is the conversation worth having, ahead of the next model cycle and ahead of the regulatory lift.
How we read the stack
Eduba’s view on AI stacks is simple. Sixty percent of most problems are plain code and a database. Thirty percent are rule logic. Ten percent are genuinely AI-shaped. Applied to Bobi:
Pulling ACOG updates and CDC bulletins into a versioned content store, with clinician sign-off on changes, is mostly pipeline and review. A small model helps at the margin. The spine is rules and human judgment.
Preeclampsia, hemorrhage, and mental-health patterns are thresholds, cohorts, and escalation rules with an AI assist. Not an LLM problem. Running this layer as rules keeps the alert path auditable and liability-safe.
Aggregating gradient updates without raw data leaving the phone is the only place in the stack where AI is doing work nothing else can do. This is the architecture worth defending and worth licensing.
Spending budget in those proportions is the difference between a six-month runway and an eighteen-month one.
Engagement shape
Feeld, a product company with a CTO trying to operationalize AI, hired Eduba for a scoped sprint: a workshop, advisory calls, an Organizational Context Architecture, and a Strategic Operations Framework their team still runs against. Their CTO engagement is the closest analog to the team inside Bobi.
For Bobi, the deliverable is that picture of the stack, plus a wedge recommendation across the three commercial paths, plus the regulatory posture plan that matches it. The output is something the team runs against, not a slide deck that dies in a shared drive.
One paper worth your time before the call
A psychometric assessment tool for evaluating ideological and moral patterns in LLMs. Maternal advice is a place where that kind of audit matters, and the BobiHealth clinical side will recognize the framing.
On the scaling question
Eduba partners with NLP Logix for work that sits below the orchestration layer: production pipelines and ML engineering at scale. NLP Logix has been in machine learning since 2011 and runs over 150 data scientists. If the sprint surfaces that layer as the next bottleneck, the handoff is warm.
Bring one workflow that eats the BobiHealth team’s week. We’ll do a live orchestration read on it, split the 60/30/10, and talk through the wedge question raised in the San Antonio Report piece.