Thinking about AI systems end-to-end. From the way a model is served, down to the silicon it runs on.
I'm Sanan. I founded Stixor to build the kind of AI systems I wanted to see in the world. Ones that actually hold up in production, at the edges of language, latency, and scale.
Seven years of this work has taught me one thing: architecture written by people who don't deploy it tends to rot. So I stay close to the code. Inference, retrieval, orchestration, data, infrastructure. I care about all five layers, and how they compose.
Along the way I've built Stixor into a 40+ engineer studio working across Pakistan, the Gulf, and Europe — the kind of team I needed around me to take AI from prototype to production.
Not an exhaustive list. The ones that best show how I think about AI in production: latency, retrieval, infrastructure, and the business around them.
Billions of transactions a month, and a rule-based system that couldn't keep up with the false-positive cost or the pace of new fraud patterns.
I designed a streaming ML system with sub-150ms inference that learns continuously from new fraud signatures, engineered for 99.95% uptime under real load.
Global LLMs don't understand Arabic legal language. Tokenisation, dialect, and juridical convention all break naïve retrieval.
A full-stack legal-AI platform built around agentic retrieval over the Saudi juridical corpus. First of its kind in the region.
HR and finance queries at telco scale were still moving through manual back-and-forth. Employees waited on expense approvals and policy answers that a system should have resolved on its own.
A conversational AI platform with a RAG layer over HR policies, finance guidelines, and operational docs — employees self-serve; answers are cited from source.
Engineers were losing real time navigating technical docs, SAP records, and field manuals spread across disconnected systems.
A generative-AI knowledge layer with hybrid retrieval over the full corpus. Natural-language questions, cited answers, SAP integration.
Standing up Pakistan's first AI-sovereign cloud — hardware, platform, GTM, governance — with no local precedent to borrow from.
I lead the hardware evaluation (Ascend NPU vs. NVIDIA GPU on performance / TCO / sovereignty), guide HCS + ModelArts on Ascend 910B, and co-develop the GTM for public-sector and enterprise.
Detecting cracks and faults in subsea pipelines and structures relied on slow, manual visual inspection.
A YOLOv5-based detection pipeline on a custom-labelled subsea dataset — full loop from annotation through training, hyperparameter tuning, and inference deployment.
One of Pakistan's largest grocery delivery ecosystems needed resilient on-prem infrastructure that wouldn't buckle under retail load.
PostgreSQL master–slave replication, Docker Swarm HA clusters, GitLab → Jenkins pipelines. Zero-downtime deploys, sane operational cost.