360Health

Built the claim settlement workflow end to end in Python/Django with a React and TypeScript front end — the claim lifecycle, the APIs behind it, and the screens each role works in.

Built the dual settlement flow so patient reimbursements and direct hospital settlements reconcile through the same path instead of two parallel processes that drift apart.

Built document digitisation pipelines that parse uploaded medical records and integrate FHIR and RxNorm, so the output is standards-shaped and searchable rather than free text sitting in a blob.

Integrated RxNorm medication APIs and wired a LangChain agent that suggests a likely insurance provider from the diagnosis on record.

Built the operational dashboards and the multi-step approval experience reviewers spend their day inside.

Worked directly with the client's engineers on domain questions, testing and agile delivery — a lot of this project was learning how claims actually work before writing anything.

Django owns the claim lifecycle and acts as the system of record. FastAPI services sit alongside it for the read-heavy paths that feed dashboards, where response time matters more than write guarantees.

A claim is modelled as an explicit state machine — initiated → provider review → hospital processing → settlement. Every transition is recorded, so the audit trail is a by-product of the design rather than something bolted on, and no screen has to guess which actions are legal next.

Uploaded documents run through digitisation and are then mapped onto FHIR resources. Medication strings are normalised through RxNorm, so the same drug written three different ways resolves to one concept and becomes searchable.

The provider recommendation is a LangChain agent given the diagnosis context. It produces a suggestion for a human reviewer to accept or reject — it never writes a decision itself.

React with Tailwind on the front end. The approval UI is one reusable multi-step primitive driven by the workflow state, not a bespoke form per role, which is what kept it maintainable as new steps appeared.

Healthcare vocabularies are unforgiving. Getting medication and diagnosis data to line up with FHIR and RxNorm meant actually learning the domain — the failure mode of guessing at field names is silently wrong data, which is the worst kind in a clinical context.

Approvals have real financial consequences, so the AI suggestion had to stay clearly advisory: surfaced with its context, trivial to override, and never applied silently. Designing for "the model is wrong" was more important than designing for when it's right.

Three roles look at the same claim and each needs a different slice of it. Most of the interface work was deciding what each role does not need to see.

Reimbursement and direct settlement look similar and behave differently. Collapsing them into one reconciliation path without losing the differences took a few iterations on the data model.

~30% reduction in claim processing bottlenecks once the end-to-end settlement workflow was in place.

~30% reduction in settlement cycle time through the dual settlement flow.

Digitised, searchable medical data measurably improved compliance and usability — records became something reviewers could work with rather than files they had to open one at a time.

LangChain-based automation cut manual effort out of the approval path.