I build software for places it never reached.

We piloted in the hardest department we could get into: paediatric haematology and oncology at Bangladesh Medical University. Children with cancer, thirty visits deep, growth curves that carry clinical meaning, and drugs that are dangerous at the wrong dose. If a record holds there it holds anywhere.

PulseAI: rules first, model last

PulseAI is the clinical assistant inside PulseNote, and it is not a language model with a prompt on top. A deterministic rules engine runs first and evaluates anthropometric z-scores and haemoglobin, platelet and absolute neutrophil count thresholds. The model is only called when no rule matches.

That ordering pays twice. Red flags, last medications and lab trends are computed from the record and returned instantly at zero model cost, which matters when the deployment is in a hospital that is not paying Western software prices. And the answers that matter most in a hurry are the ones no model touched.

Every question and every answer is written into that patient's audit log. If PulseAI told a clinician something, there is a record of it that cannot be quietly edited later.

The answers that matter most in a hurry are the ones no model touched.

Dosing is arithmetic, not generation

The dose engine is a pure function. Mosteller body surface area, then weight based, BSA based, fixed and volume per kilogram dosing, with hard caps at adult maximum single and daily dose. If a high alert drug is missing its safety bounds the engine refuses rather than estimating, and high alert drugs additionally require a second clinician to confirm, which is written to the audit log.

The short read, and why it has to cite

A summary spans up to thirty past visits and turns years of labs, medication changes and growth data into something readable in the time a consultation actually lasts. It runs on two separate search indexes: an institutional knowledge base the doctors curate themselves, and the patient's own record. Every factual statement must carry its source, and the model is barred from stating a drug dose that is not traceable to a retrieved document. A summary a doctor cannot check is worse than no summary, because it is faster to trust.

Privacy by design, not privacy by policy

A privacy policy is a promise. I wanted the guarantees to be structural, so that doing the wrong thing requires going out of your way rather than merely forgetting. Consent is granular and it belongs to the patient, three tiers, live on web and on mobile. Roles are resolved server side, never from anything the client can edit. And there is exactly one path to a full patient record: a single authorization gated server side join. One door, so consent, purpose of use and audit happen in one place or they do not happen at all.

Two details that matter more than they sound. Errors from upstream are redacted before they are returned, because a stack trace is a data leak wearing a different hat. And access is written to a hash chained ledger, so the log can be verified rather than trusted, which is the difference between an audit trail and a list.

Building for a standard nobody is asking me for yet

Nobody in Bangladesh is currently asking for FHIR. I built the export pipeline anyway. The problem I set out to solve was one patient's history being unreadable, but a record that only works inside my system recreates the same trap one layer up, with the paper folder replaced by a vendor. So the internal model maps to FHIR R4 across Patient, Condition, Encounter, Observation and MedicationRequest, behind an authenticated, rate limited endpoint with 55 tests gating every deploy, plus an inbound HL7 v2 path sharing the same mappers.

Honest scope: server side API, no clinician facing export screen yet, not run against an official conformance suite. What exists is the pipeline, so the day a record has to move to another hospital, vendor or country, that is a configuration problem and not a rebuild. A record that cannot leave the system was never really the patient's record.

The first version could tell a doctor a child looked fine. I removed that result. The department's own study of 223 children with childhood leukaemia found that some confirmed cases arrived with a blood count that looked unremarkable. So the tool has no way to say all clear. It flags a case for a doctor to look at, or it stays silent. It never reassures.

I use AI heavily and I say so. Where an AI generated document disagreed with the department head's own instructions, the instructions won. The AI does not get to make the clinical call.

Same failure as the folder, different room. The information existed. It was just held in a form nobody could act on in time.

Travista OS runs the whole business now. Accounting, inventory, sales, corporate clients, all in one system. The things that used to eat a day, doing the books and knowing what was in stock, got easy, and the team gets more done in less time.

The part I am proudest of is who uses it. Nobody on the floor is technical. They open the Android app, scan an item and see its details, watch the day's sales, look after a corporate client, all from a phone. What took a ledger and a phone call now takes a scan.

It works because I built what they actually needed, not what I wanted to build. I was in the shop to see what that was.

I built what they needed, not what I wanted to build.

Stock that cannot half happen

Goods receipt is one database transaction. All of it commits or none of it does, so stock never moves without something recording that it did.