I build production AI for KSA payers and providers — pre-authorization, fraud-waste-abuse, clinical triage, and automated medical coding — engineered around NPHIES, SBS V3, and Arabic from day one. Deterministic where it must be. LLM-powered where it helps. Never LLM-only on a medical decision.
Each system is in production-grade engineering: typed code, automated tests, observability, and audit trails. No demoware.
A four-layer decision pipeline — health declaration check, deterministic rules engine, RAG-retrieved policy interpretation, and a medical reasoning model. Outputs an instant, explained authorization decision with the policy clauses and clinical evidence cited inline.
Real-time dashboards for payer pre-auth teams — TAT gauges, hourly throughput, HCP heat-maps, processor productivity and approval/rejection ratios, SLA-breach alerts and escalation, scheduled email digests for managers, region directors, and senior leadership. Ramadan-aware scheduling included.
Retrospective review of pre-auths and claims against CCHI regulations and payer-specific policy. A versioned rules engine catches the known. ML anomaly detection (isolation forests, duplicate/upcoding/unbundling detectors with peer-group benchmarking) catches what the rules can't. LLM enrichment runs only through a zero-trust PHI-masking pipeline that fails closed on any leak.
Bilingual triage and home-program app for the four most common musculoskeletal regions, with browser-based pose estimation that counts reps and scores exercise form in real time. Red-flag rules short-circuit immediately to urgent care. Positioned as triage and education — deliberately below the medical-device line, with regulatory review tracked openly.
Maps internal hospital and payer codes to the Saudi Billing System V3 dictionary mandated by the Council of Health Insurance. Three-stage pipeline: clinical-domain embeddings, vector similarity, then LLM clinical-meaning reasoning, then a 37-rule validator. Every match comes with a confidence score, the candidate alternatives, and a reviewable audit trail.
Rules engines, schemas, and version-pinned policy documents own the decision. The LLM extracts, summarizes, and proposes — it does not authorize. Two identical inputs always produce the same output.
No protected health information leaves your perimeter to a third-party model without policy-driven masking — and the pipeline halts if a single masker fails. Every model call is logged with input hash, output hash, and the masking version that ran.
Arabic free-text input, RTL-correct interfaces, Hijri date awareness in operational schedules, and ICD-10 ↔ NAPHIES ↔ SBS V3 mapping at the data layer. Arabic isn't a translation pass — it's a first-class user.
I'm Dr. Mohamed Gohary — physician by training, engineer by practice. I started in clinical work, saw firsthand where time, money, and patient outcomes leak out of the system, and moved into building the software that closes those gaps. Today I design and ship AI systems that real teams use every day across KSA pre-authorization, claims integrity, clinical triage, and medical coding.
What I bring that most vendors don't: I have done the workflow. I have sat with the processors at 2 a.m. during Ramadan night-shift TAT spikes. I have read the policies and argued the clinical exceptions. The result is software that feels like it was built by someone on your team — because the person who wrote it has been on that team.
If it's healthcare in KSA and it touches data, policy, or workflow — I want to hear it. Most replies within one working day.