Software engineer. I build things end to end.
I'm Yassine Belghiti, a software engineer based in Morocco. I build full-stack products and AI systems — retrieval pipelines, voice agents, back-office automation — and I lead the engineering team at LeetBridge.
I trained at 1337 (part of the 42 network) and through UM6P's programs, where I also did hands-on engineering on university systems. Most of what I build starts as a real problem someone I know is stuck on, and I stay close to the code on every project.
A few projects I've built end to end. Most are production systems; a couple are tools I use myself. Each one links to a short case study with more context.
A voice agent that calls customers to confirm cash-on-delivery orders, replacing a manual call-center step for e-commerce sellers. It handles the whole conversation — speech, understanding, response — at low enough latency to feel like a normal call.
Python, FastAPI, Twilio, Whisper, ElevenLabs, OpenAI, Redis
Turns raw chess-engine output and opening theory into plain-language coaching. A retrieval pipeline pulls the relevant theory for a position, and the model explains the idea instead of just naming the best move.
Python, FastAPI, Stockfish, pgvector, OpenAI, React, GSAP
A compliance and productivity platform for Moroccan accounting firms and their small-business (TPME) clients: DGI-compliant electronic invoicing, automated tax and regulatory filings, and a client alert system, on top of a documented API and a clean interface.
Django, PostgreSQL, Celery, React, Redis
A personal knowledge system with modules for learning frameworks, execution checklists, a Moroccan business-law reference for company structuring, and a sector-intelligence module covering fintech, greentech, agritech, and retail tech.
Next.js, TypeScript, Supabase, OpenAI, Vercel
Not shipped yet — things I'm actively working on. Listed here so the status stays honest.
A Model Context Protocol server that exposes my projects, CV, and case studies to AI clients like Claude. A working MCP demo, not just a description of one.
Specialized agents — research, code, critique — working on one task, with a dashboard that shows the handoffs and reasoning as they happen, not just the final output.
Retrieval that combines a knowledge graph with vector search for multi-hop questions, instead of relying on vector similarity alone.
A DSPy-style framework that tests prompts against an example set and improves them iteratively, with before/after metrics.
Traces, scores, and flags regressions — hallucination, latency, cost — in LLM pipelines. The production-readiness layer most AI demos skip.
I start from the problem, not the stack. The useful question is usually what someone is doing by hand and why, not which framework to use.
I build the whole thing — data, backend, model, and the interface someone actually clicks. The joints between those are where projects break.
I ship something small that works, then extend it. A pipeline that runs today beats an architecture diagram.
I've played chess seriously for a long time. It mostly taught me to plan a few moves ahead and to play the position in front of me instead of the one I wish I had.
I trained at 1337, the Moroccan campus of the 42 network — peer-to-peer, project-based, no lectures. I went through UM6P's entrepreneurship and innovation programs, and did engineering work inside the UM6P ecosystem (GEPS, SIL).
Ventures I'm involved in
The best way to reach me is the form below — technical work, advice, or just to compare notes.