London Area, United Kingdom
more info at: mohamedsharif.tech My goal is to help build something millions of people use and love. Think Notion, Figma, Claude. The opposite is imagine your workplace and you’re forced to use the most outdated, cumbersome, hateful app for your payroll, or ticketing or communication. That’s where I want to stay as far away from
Forward Deployed AI Engineer, Building Sovereign AI for Government
Now building Proposer - an AI Legal Mediation System. It uses a hybrid RAG + Knowledge Graph architecture to predict tribunal outcomes and facilitate fair settlements, no lawyers required. With Glass-Box reasoning to understand fully why a judge might land on a settlement, as well as rational mediation to help propose satisfiable outcomes for both landlords and tenants. Prev: Building an AI memory layer integrated with your browser so that when you use your favourite chatbot, it has automatic context from the content you've researched. (Parked)
Engaged with prospective students from diverse backgrounds, promoting STEM fields and inspiring the new scientific brightest minds during summer schools and open days. Conducted academic sessions as well as university support sessions in various different areas and subjects. Presented at educational institutions as well as honed essential skills in communication, teamwork, organization, and marketing.
Worked on large-scale AI training workloads at Amazon Web Services (between S3 and Pytorch) where I optimised the AWS S3 Connector for PyTorch to speed up distributed ML training jobs. Benchmarked performance using Python/Boto3, analysing latency & throughput to guide design improvements. Scaled distributed checkpointing to support workloads from 80 instances to 300+ EC2 instances (1M+ S3 requests/min). Built a “shadow copy” system that reduced S3 5xx errors by 90%, making training pipelines more reliable. Added error-rate analysis + logging dashboards to pinpoint I/O bottlenecks in large training jobs. Analysed model-training patterns and FSDP sharding behaviour from AI Labs to improve data-loading efficiency and training reliability.
1.3k community members >30 events > few thousand attendees organized 2 24-hour hackathons ...nearly burnt out