Mohamed
Yousry
I build production AI systems, scalable backends, and automation that actually ships. I run models on hardware they weren't meant for, put most of what I build out in the open source, and care a little too much about systems that stay up. Right now I'm leading engineering at Drive-Nova.
Where I've worked
Lead Engineer
Own the architecture and delivery across B2B and B2C platforms in the tyre industry.
- Lead engineering across B2B and B2C platforms, including managed Magento storefronts, owning the architecture, the delivery, and keeping it all running.
- Architecting the AI side of the roadmap: agent-to-agent (A2A) architecture and multi-agent frameworks like AG2.
- Run integrations and the technical relationships with partner platforms across Europe, Egypt, and the UAE.
Fractional CTO & Lead AI / Backend Engineer
Built a full-stack AI procurement platform from the architecture up to production.
- Designed and shipped an AI procurement platform that aggregates 50,000+ federal (SAM.gov) and SLED (HigherGov) opportunities, with multi-schema PostgreSQL, semantic search over embeddings, and an event-driven processing system.
- Built ELI, a conversational assistant with 24 procurement tools (opportunity search, document Q&A, board management) on GPT-5 / GPT-4o, plus a 4-tier matching pipeline and a PaddleOCR pass for scanned government forms.
- Architected the microservices: FastAPI, Celery with 6 specialized workers, Redis, and Supabase. Shipped the whole production stack on Docker, Coolify, Nginx, Cloudflare, GitHub Actions, and Sentry.
AI Engineer & Automation Developer
Owned payments and infrastructure for an AI marketing product.
- Built and maintained the Stripe stack: subscriptions, webhooks, and secure checkout.
- Wired frontend flows into the backend APIs and tightened up performance along the way.
- Set up and ran the infrastructure: domains, DNS, Nginx, and Git workflows.
Co-founder & AI Engineer
Co-founded a company building local, on-device AI.
- Led the AI features for a local, on-device AI product built with Flutter.
- Designed and deployed the scalable backend on FastAPI, PostgreSQL, and Docker.
- Set up the CI/CD pipelines that kept deploys boring (which is the goal).
Developer
The long-running thread. Shipping production systems end to end since 2020.
- Architected and deployed large systems end to end: AI pipelines, OCR systems, mobile apps, web platforms, and automation bots.
- Built OCR and document-processing systems that handle preprocessing, extraction, validation, and clean structured output.
- Ran the full infrastructure lifecycle: cloud setup, provisioning, Dockerized deploys, CI/CD, DNS, and production monitoring.
Stuff I've built
The setup that actually gets PyTorch and Unsloth fine-tuning running on an AMD Strix Halo (gfx1151) with 128GB of unified memory. I tracked down the kernel regression and the exact ROCm versions that quietly break every other guide, then shipped prebuilt containers so nobody else loses the weekend.
A self-hosted Drive where the server only ever holds ciphertext. Your keys are derived in the browser with Argon2id, sharing seals a file key to the recipient's X25519 key, and nothing sensitive leaves your machine in the clear. Dump the database and all you get is base64.
A multi-agent ops platform where a LangGraph DAG runs department agents over an async agent-to-agent message bus and streams the whole thing to a React dashboard over SSE. This is the A2A architecture from my resume, as working code you can read.
A privacy-first grammar checker and rewriter in five languages. It runs LanguageTool's rules first, then an LLM, then validates the LLM against itself to kill the errors AI usually introduces. Flutter app, FastAPI backend, encrypted end to end.
An IDE driven by 11 orchestrated MCP agents for coding, debugging, testing, and deploy, plus a free community marketplace for tools and extensions. Local-first by design: SQLite and ONNX embeddings run right inside Electron, with optional Supabase sync.
A multi-agent tool that scores research papers against the COSMIN risk-of-bias checklist, all ten boxes, with verbatim evidence quotes you can click straight to the right page in the source PDF. Built for real systematic reviews, not demos.
An Instagram Reels pipeline that downloads in full quality, strips metadata so it stays under the radar, transcribes with Whisper, ships to S3, and runs the whole flow through n8n. FastAPI backend, one-command Docker deploy.
Pairs ML clustering with Llama 3.2 or Gemini to read patient data and surface severity, estimated cost, and triage priority. Runs fully local with Ollama or in the cloud, your call.
Turns any image into a numbered pixel-art pattern for cross-stitch, Perler beads, or LEGO. Color clustering, optional dithering, multiple export formats. The fun one.
What I actually use
Off the clock
I run ML on hardware it was never meant for, then write down what finally worked so the next person doesn't lose a weekend.
I'm an open-source person at heart. Most of what I build ends up public, and I lean on the community's work every single day.
I'd rather read the logs than guess. Most 'impossible' bugs turn out to be something boring I just hadn't looked at yet.
Off the clock it's music production and video editing. I edit better than I film, and yes, I can burn garlic in under two seconds.
Say hi
Let's build something that stays up.
The fastest way to reach me is email. I read everything and reply to most of it. Got a hard backend or AI problem? Even better.