Insights · 20 articles · 5 topics

How this softwareactually gets built.

Notes from the people building hospital platforms, clinical AI and the integrations between them. Written for the engineers, clinicians and security teams who have to live with the result.

No thought-leadership theatre. What we decided, why, and what we would do differently.

FeaturedInteroperabilityOne record, twelve systems: what a patient journey really touchesA single visit passes through registration, the ward, the lab, pharmacy, billing and discharge each often its own system. We walk through where the continuous record breaks today, and the design decisions in Zypocare One that keep it whole from access to follow-up.Zyposoft Engineering21 Jul 20265 min readRead article

Reading

Latest articles.

Newest first. Filter by what you work on.

Responsible AIDrift is the failure mode you will not noticeA model that fails on day one gets fixed. One that degrades slowly as the population, the coding practice or the upstream data changes will keep producing plausible output while quietly becoming wrong. What we monitor, and why the alerting points at inputs rather than accuracy.17 Jul 20263 min read Responsible AIAI that hands you something to check, not something to trustA model that cannot explain itself has no business near a bedside. How Zypo Clinical AI is built so a clinician gets a suggestion with the reasoning attached and keeps the decision, the override and the audit trail.14 Jul 20263 min read InteroperabilityThe identity problem nobody budgets forMatching a person to their record is the hardest problem in health integration and is almost always estimated as a checkbox. Duplicates, merges, near-matches, and why the auto-match threshold is a clinical decision rather than a technical one.10 Jul 20263 min read InteroperabilityHL7 and FHIR in practice: connecting the systems you are keepingNobody replaces everything at once. A field guide to the message types, mapping traps and versioning realities of wiring a new platform into the EMR, LIS and PACS a hospital already runs without a rip-and-replace.7 Jul 20263 min read SecurityAudit trails that answer questionsMost audit logs are written to satisfy a requirement and discovered to be useless the first time somebody asks a real question of them. What changes when the log is designed around the questions instead.3 Jul 20263 min read SecuritySaying 'encrypted' precisely'Encrypted' on its own is marketing. What it takes to make it a statement a security reviewer can check: the algorithms, the parameters, and the discipline of describing exactly what each one protects.30 Jun 20263 min read Responsible AIWhat we ask before a model goes near a patientGoverning clinical AI is mostly a set of questions asked early enough to still change the answer. The ones we work through before anything is built, and the ones that have stopped features.26 Jun 20263 min read HealthcareDesigning for three in the morningSoftware in a hospital is clinical equipment. It gets used by someone who has been awake for fourteen hours, not by someone at a demo. What that single constraint changes about interface, error messages and defaults.23 Jun 20263 min read EngineeringChoosing boring technology on purposeIn a domain where software outlives the team that wrote it, novelty is a liability with a delayed invoice. How we decide when something new is worth it, and the questions that usually settle it.19 Jun 20263 min read HealthcareBuilding for ABDM: ABHA, consent and national registriesIndia's digital health stack is not a compliance chore bolted on at the end. How ABHA identity, consent-driven exchange and standards-based records fit into everyday hospital workflows rather than sitting beside them.16 Jun 20263 min read HealthcareThe ward round is the specYou cannot design clinical software from a requirements document, because the document describes the process as somebody believes it works. Watching a ward round for two hours changes what you build.12 Jun 20263 min read EngineeringThe unglamorous release: shipping without a drama every timeAnyone can stop a release here. Why 'correct beats fast' is a process and not a slogan, and the review, migration and rollback habits that let a small team put clinical software live at a hospital and sleep afterwards.9 Jun 20263 min read InteroperabilityYour integration layer is a product, not a projectInterfaces get funded as one-off builds and then run for a decade with no owner. What changes when the integration layer is treated as something with a roadmap, a support model and a person whose name is on it.5 Jun 20263 min read SecurityWho can see what: identity and access before anything elseConsent, roles and context are settled early or patched forever. A look at how identity, RBAC and context-aware access are treated as part of the design in Zypocare One, not a layer added once the product works.2 Jun 20263 min read HealthcareWhat a discharge summary is actually forIt is the only part of an admission most of the outside world will ever read, and it is usually written last, fastest, and by the person with the least context. A look at the document as a design problem.29 May 20263 min read EngineeringSmall teams, real ownership: how enterprise product work runsThe same engineering that goes into our hospital platforms goes into enterprise product work: named owners, decisions written down, and a short line between the person who decides and the person who lives with it.26 May 20263 min read EngineeringThe cost of a feature is not the buildEstimates cover the construction and ignore the decade of ownership that follows. What we try to count instead, and why the cheapest feature is frequently the one that was not built.22 May 20263 min read SecurityThe risk you inherit from everything you connect toA platform's security posture includes every system it integrates with and every credential it holds on somebody's behalf. How we think about the blast radius of an integration, and why least privilege is an integration requirement.19 May 20263 min read Responsible AIModel documentation a clinician will actually readTechnical model cards are written for reviewers and read by nobody at the bedside. What a one-page description has to contain for the person deciding whether to act on an output.15 May 20263 min read

Stay in the loop

Get new writing when it lands.

Occasional, technical, and never a newsletter for the sake of one. We will only write when we have something worth your time.

Email us to subscribeNew writing, now and then Talk to us