How Compiled Health uses Devin to make the standard of care executable

The two-minute problem

Today, Compiled's preventive-care modules are live across Canadian provinces. Across roughly 4,000 patients, they've delivered nearly 16,000 recommendations. The team is about four people.

In a primary care visit, the doctor is focused on whatever brought the patient in: an asthma attack, a mental-health concern, or another immediate issue. Preventive care often gets pushed to the last couple of minutes.

But figuring out which screenings a patient needs is not a simple checklist. Each recommendation depends on age, risk factors, prior results, potential harms, and the specific clinical guideline. Those guidelines can run hundreds of pages and change over time.

Compiled turns that guidance into deterministic software that runs during the visit. Instead of relying on a physician to remember every applicable rule, the system produces a complete care agenda and shows the logic behind each recommendation.

Making clinical guidelines executable

Compiled uses AI while building its clinical modules, but not at runtime in front of a patient. The team extracts the logic from source guidelines, resolves conflicts, and turns it into versioned clinical decision modules. Each release records what changed, when it changed, and which version governed a patient's assessment.

That distinction matters in medicine. A physician needs to be able to inspect why a recommendation was made and reproduce the same result from the same inputs. Compiled's goal is deterministic clinical logic that can be traced back to the source guideline.

Why they chose Devin

The core engineering problem was much bigger than extracting rules from PDFs. Compiled also needed to prove that the resulting logic worked across the full range of possible patient inputs.

A single guideline can apply to millions of combinations of patient features. A small team of mostly physicians could not test those combinations by hand, and Compiled did not want to build a large engineering organization before it could prove the product.

Christian compared coding-agent harnesses head to head in a months-long pilot. He wanted an agent he could push beyond a model's default clinical intuition and use across medicine, mathematics, and software engineering. Devin stood out for its flexibility, its ability to carry context across disjoint ideas, and its ability to take work from an initial idea through a tested implementation.

Proving the logic correct

The hardest part was verification: proving that a compiled guideline had not omitted a case, introduced a contradiction, or created an unreachable branch.

Devin surfaced mathematical research on decision-tree verification and helped implement a verification system for Compiled's guideline library. The checks test completeness and internal consistency across the decision logic. The alternative would have been thousands of expert-physician hours spent stress-testing modules manually.

For Compiled, this was foundational to the product.

"Verifiability was never optional for us. Without it there's no product, much less a case to be made for reliability. The question was whether four people could build this in a year or in five. Devin collapsed timelines to months."

— Christian Chartier

Building a new clinical benchmark

Compiled is now using the same approach to build a clinical benchmark it plans to release this fall. Existing medical benchmarks often test whether a model can answer medical questions or pass an exam. Compiled's benchmark asks a different question: does the model actually follow the clinical standard, and did it collect the information required to reach its recommendation?

Christian estimates the project would have required five or six engineers working for many months, plus an unusual combination of medical and mathematical expertise. Instead, two researchers brought it near completion in two months with Devin augmenting their engineering capacity.

The broader pace is similar. In one recent sprint, the team shipped in 17 days roughly what it had shipped in the previous nine weeks. Over a two-week period, Devin delivered around 430 measured hours of engineering work at roughly $17 per hour — equivalent to adding more than four engineers to a four-person company.

Cognition-native engineering

Compiled describes itself as Cognition-native: every new hire is onboarded to Cognition on day one. But the team does not ship Devin's work on trust. The same verification mindset that underpins the clinical product also shapes how the team uses Devin.

The relationship extends beyond the product itself. After reading Cognition's research on agentic map-reduce, Christian adapted the technique inside Compiled's system to check guidelines against thousands of synthetic patients. The teams also trade ideas directly, with Compiled pushing Devin into a domain where software correctness is a clinical requirement.

The road ahead

Compiled is building toward a world where every doctor can apply the full standard of care during a normal patient visit, backed by clinical logic that is deterministic, inspectable, and verified.

That is normally the kind of problem that takes large teams and long institutional timelines. Compiled is trying to do it with four people.

"Devin takes something really hard and just slices through that difficulty like butter."

— Christian Chartier