How we helped a medical-device company avoid a $14M+ study with a model delivered in 5 weeks.

They needed one missing model to secure their place in the study.

The insulin-pump company wanted its system included in Novo Nordisk's feasibility study of a new insulin formulation, but taking part meant showing that its dosing algorithm could handle the new insulin before the study deadline.

That required a model of how the insulin entered the body and how its glucose-lowering effect changed over time. The company's outside modeling advisor had retired ~6 months earlier, and its search for a replacement hadn't succeeded, leaving the internal team to work on the missing model.

After several weeks, the team had reached ~52% first-hour accuracy against an acceptance standard of 72%, which its production system already met for Novolog, the insulin it supported.

With the deadline approaching, that gap put study participation at risk and left a costly alternative: funding a separate study that leadership estimated would cost more than $14M.

We confirmed the missing capability before putting the limited time into a build.

The company came to us through a referral with a clear diagnosis, and we spent a few days examining its model, the dosing algorithm's handling of Novolog, and the evidence it needed for study participation.

What we found confirmed the team's assessment: the dosing algorithm already worked with the existing insulin, but the company needed a model of the new formulation that met the same standard. We agreed to build and evaluate that model while its algorithm team handled simulator integration and the dosing algorithm itself, bringing in the missing specialist capability without moving responsibility for the device outside the company.

An eight-day data delay made the order of work matter.

Both partner contacts were away at the same time and no backup had been assigned, which meant usable data arrived eight days into a five-week engagement.

We could still prepare the work needed to judge the model, and so we brought the acceptance conditions, fitting pipeline, and evaluation tools forward. The company already had a known Novolog reference that we could use to make the evaluation reproduce a result its team could verify, so the tests were ready when the new data arrived.

We kept the client informed about what was progressing while we waited, but the delay also exposed a gap in how we'd planned for outside inputs.

The model exceeded the agreed standard, and the team could reproduce the result.

When the data arrived, we used it to build a model of the new insulin's absorption and effect, drawing on information across the available participants while accounting for individual differences.

We compared the model with Novo Nordisk's clinical reference data across the combined profile, individual curves, and subgroups. First-hour model accuracy reached ~86% against the agreed 72% standard, up from ~52% in the company's earlier attempt, giving the team evidence that the missing model met the bar it had set.

That figure measures agreement with the reference for the insulin's effect over the first hour, rather than patients' time in a target glucose range.

We should have secured a backup for the data dependency.

We kept the schedule moving by bringing the evaluation work forward, but we'd entered the project without a named backup or escalation path for a critical partner input. On a project with less independent work available, the same eight-day wait could have pushed delivery beyond the deadline.

That planning gap was ours, and it changed how we prepare for work that depends on another organization. Our kickoff planning now identifies each outside input, its owner and backup, the escalation route, and what can continue if the input arrives late.

Their team reproduced the result before accepting it.

Clearing our own test wasn't enough for their team to accept the model, and so we handed over detailed methods and evaluation documentation that their algorithm team used to reproduce the result on its own machines.

Acceptance moved through their VP of Engineering and CTO/CPO, then to Novo Nordisk, with reviewers able to compare the deliverable against conditions agreed before the model existed. That verification resolved the final open requirement for participation, allowing the company to secure its place in the feasibility study inside the deadline and avoid the separate study route.

Their team then took the evaluated model into its own integration and device-testing work. Study participation was a program milestone, with regulatory approval of a device still a separate matter.

For a closer look at the model and its evaluation.

The technical companion explains what the insulin model measures, how we evaluated it against the reference data, and how the company's team reproduced the result. It also covers the data delay, acceptance process, and scope of the study milestone.

They secured study participation without funding the costly alternative.

The company needed a model that met an existing standard before a fixed deadline, and we confirmed that requirement before building and testing it in five weeks. We gave its team the documentation to reproduce the result, which resolved the final open requirement for study participation and let the company secure its place on time.

If you’re trying to grow your own business, start with your goal, look at what it'll take to achieve it, and then identify what's keeping you from taking action. Just remember that every business is different and that a missing specialist model was just their bottleneck and it might not be yours.