FlashKernel
CUDA and Triton kernels for attention, activation fusion and cache operations. Explore the implementation, correctness boundaries and GPU validation requirements.
Explore FlashKernel →
Abdeljalil Jliouat
AI engineering · Paris
Founder of AKIOUD AI and Lead AI Engineer at Orange, where I work on agentic applications, RAG and LLMOps for Finance and CFO teams.
A closer look · RoboLLM
Can the robot approach without disturbing the scene?
A fixed goal above the object, a full second of holding, and checks on collisions and object movement.
- Method
- Two controllers, identical initial states across 100 fresh simulated scenes, and checks every 2 milliseconds.
- Finding
- 88/100 successful holds with a collision-aware plan, versus 29/100 with direct control. The 12 cases without an accepted plan count as failures.
Open projects & experiments
Code and experiments in inference, robotics, neural signals and optimization. Each project documents its current scope and limitations.
RoboLLM
A MuJoCo prototype tested on fixed-goal reaching and one-second holds, with collision checks, complete traces and reproducible controller comparisons.
Explore RoboLLM →NeuroLLM
An EEG transformer prototype for masked reconstruction and motor-imagery classification, with data separation and evaluation as the current priorities.
Explore NeuroLLM →QuantumGrid
Synthetic unit-commitment experiments using QUBO, QAOA, VQE and classical solvers. Explore the formulation and requirements for a fair comparison.
Explore QuantumGrid →Experience
Leading applied ML and AI systems across finance, banking, startups, and NGOs.
-
Founder @AKIOUD AI, Paris–Present
-
Lead AI Engineer · Orange–Present
-
Lead Data Scientist, Merci Jack, Paris–
-
Lead Data Scientist, Société Générale, Paris–
Have a technical problem in mind?
Tell me what you’re building, where it gets difficult and the constraints that matter.