Case study · 2024 — Present

Precision Health Platform

A clinical genomics platform built from an empty repository to production — a million classified variants, ACMG classification, polygenic risk scores, and a physician-configurable AI layer on top.

Next.js
FastAPI
PostgreSQL
Genomics
ACMG
PRS
Pharmacogenomics
Docker
RBAC
HIPAA

Repository

The problem

The largest system I have built end to end. A patient registers, completes intake, then uploads whatever they have — blood work, a microbiome panel, a raw 23andMe or AncestryDNA export — and the platform keeps analysing it rather than producing one report and stopping.

Underneath is a genomic analysis engine holding close to a million classified genetic variants, and above it a personalised assistant that a physician can configure for their own patients.

What it does

Raw genotype files classified against ACMG guidelines, cross-referenced against ClinVar and GWAS Catalog.
Polygenic risk scores across a few hundred conditions, alongside drug-gene interaction analysis for pharmacogenomics.
Multi-omic integration: genetics, clinical chemistry, microbiome, mitochondrial function, epigenetic age and wearable data in one model of the patient.
Parsers for dense lab formats including MITOswab and GI Effects, with output rendered in English or Persian.
An ensemble AI layer — several models where that improved reliability, one model where it did not — plus a peptide and pathway analysis layer.
Role-based access control and audit logging across admin, doctor, technician and patient roles.
Interactive 2D and 3D visualisations mapping findings onto the body, and exportable patient-ready clinical reports.
Containerised and cloud-deployed, designed against HIPAA and GDPR from the start.

Health Hub showing the overall health score and eight data sources — Fig. 1 — One score built from eight independent data sources — genomics, risk scores, blood labs, microbiome, epigenetics, mitochondria, wearables.

SNP database browser showing 998,204 genetic variants by category — Fig. 2 — 998,204 classified variants, grouped into the categories the clinical rules query against.

Dashboard with clinical findings and genetic risk distribution — Fig. 3 — Clinical findings ranked critical before warning, with the polygenic scores that produced them.

Agent Builder screen for configuring a custom AI advisor — Fig. 4 — The Agent Builder — a physician configures a specialised advisor, scoped to the data it is allowed to read.

What was hard

A million variants is a data problem before it is an ML problem

Classification is only as good as the reference set behind it. Most of the work was building and reconciling the variant database — ClinVar and GWAS Catalog disagree, evidence levels change, and a stale classification in a clinical context is worse than no classification.

Language models hallucinate confidently about genetics

The AI layer never invents a finding. It is constrained to interpret data the engine has already classified, which is why the ensemble exists — agreement between models is used as a reliability signal, not as a way to sound more certain.

Continuous analysis, not a one-time report

Every new upload has to re-open conclusions drawn from earlier ones. That reframes the whole system from a report generator into something closer to a longitudinal record with an opinion.

Other case studies

(2)