YosriKhiari
Optimising: what holds under load • Not: what demos well
khiariyosri02@gmail.com · linkedin.com/in/yosrikhiari · github.com/yosrikhiari
Build it. Then make it hold.
Systems that survive traffic, tenants and time — not just the demo.
I'm a software engineer who ships production systems across the full stack — from AI pipelines and retrieval-augmented generation to multi-tenant SaaS platforms and container orchestration. I care about the part after the model returns: how the request is classified, retrieved, filtered, traced, and kept honest.
Over thirteen months of internships across three companies I've led the development of a semantic search engine (6 custom services on a 29-container stack), owned ERP modules from database to UI, built ML forecasting pipelines, and run Kubernetes deployments across environments. My toolbox spans Python, TypeScript, C#/.NET, Go and Java.
I'm finishing an Engineering Degree in Computer Science at ESPRIT in Tunis (September 2026), after preparatory studies in Mathematics and Theoretical Physics at ISSAT Mahdia (ISSATMA) — which is where I learned to ask why a thing works before trusting that it does.
The trajectory.
Systems that had to hold.
Side projects built to production standards — tests, monitoring, and load included.
Versatile
Offline-first AI fiction writing platform. A LangGraph.js graph runs Writer, Critic and Editor as separate agents on GPU and CPU lanes — the critic judges one scene while the writer drafts the next — checkpointed per step, with a live Agents panel and every model call traceable through AgentOps. Vue 3 + .NET 10, Dexie/IndexedDB sync, TipTap editor, five AI providers. ~3,100 automated tests.
AgentOps
Production-grade benchmarking platform in Go. Stress-tests LLM providers under realistic traffic with Prometheus/Grafana monitoring, k6 load tests and 11 configurable policies — rate limiting, circuit breakers, cost caps.
Chexy
Chess platform with a classic mode and an RPG mode (armies, gold, abilities). Spring Boot API with Keycloak auth, MongoDB and Kafka game events; React client over STOMP WebSockets; a Flask service that asks Stockfish for moves and scales them to the bot’s skill. Deployed to Kubernetes by a Jenkins pipeline. Now one monorepo with the four parts’ full history.
Jester
Web intelligence pipeline. A Go worker scrapes 80+ community sources, a chain of LLM agents (extractor, synthesizer, critic) distils pain points, and a React 19 dashboard surfaces scored product ideas.
What I build with.
Seven clusters, one graph. Tools I've actually shipped or trained with — hover a hub to light up its cluster, drag anything.
Hard problems welcome.
Looking for a team where the interesting part is everything around the model. A question, a role, a project — the inbox is open.