Open to internships · from Dec 18, 2026

I build AI from the paper up.
And ship it to real users.

Computer science student at CESI, Paris. Six years training models straight out of research papers, and two internships building the AI backends behind products used by thousands.

01 The stack

Every layer.
From silicon to screen.

Most engineers pick one layer. I work across all of them — reading the paper, training the model, squeezing the GPU, and shipping the backend people use on day one.

Scroll to unfold

02 Experience

Shipped to production.
Twice.

01 Apr — Jul 2026 Algeria · Internship

ISR Services

Backend / AI Engineer

~2,000users within weeks of launch
  • Built the backend of wandit.dev, an AI platform that turns a single prompt into websites, landing pages and Meta / TikTok ad campaigns.
  • Orchestrated LLMs to generate ad visuals, promo videos and complete Meta / TikTok campaigns.
  • Took it from zero to roughly 2,000 users a few weeks after launch.
LLM orchestrationNode.jsGenerative mediaAds APIs
02 Sep — Oct 2025 Paris · Internship

VEROOM — GoGain

Backend / AI Engineer

iOS + Androidlive on both stores
  • Built the AI backend of GoGain: personalised meal plans and training programmes.
  • Designed and orchestrated LangChain agents, cut response latency, and integrated a medical / fitness RAG with conversational memory for contextual recommendations.
LangChainRAGAgentsFastAPI
0years training models from research papers
0users on Wandit within weeks of launch
0parameter DeepSeek-style MoE, built from scratch
0faster training on my Tiny Recursive Model rework

03 Projects

Papers in.
Working code out.

I learn an idea by rebuilding it. These are a few of the models and products I've written from the ground up.

01 · Architecture

DeepSeek-V3-MoE

A ~82M-parameter DeepSeek-V3 rebuilt from scratch: Multi-head Latent Attention to shrink the KV cache, 1 shared + 8 routed experts with top-2 routing, and the Muon optimizer. Pre-trained on TinyStories.

PyTorchMLAMoEMuon
View code ↗

02 · Reasoning

Improved Tiny Recursive Model

A re-engineered TRM for iterative reasoning with very few parameters. I fixed broken backprop-through-time and fused kernels: ~4.5× faster training, ~5× faster inference.

RecursionBPTTKernel fusion
View code ↗

03 · Reinforcement learning

Reinforcement-Learning

Value- and policy-based agents from scratch — Q-learning, DQN, Double / Dueling DQN, PPO, actor-critic, GAE. They play Pong, land on the Moon, learn to walk, and fight in VizDoom.

PPODQNGAEGym
View code ↗

04 · Research

Papers-in-code

Research papers implemented from scratch: Transformers, RNNs, LSTMs, Gemma 2 / 3, Vision Transformer (DyT) and VAE — with AMP and gradient checkpointing to keep memory in check.

TransformersGemmaViTVAE
View code ↗

05 · Product

Wandit

An AI platform for web and marketing. I built the backend, the LLM orchestration and the integrations that generate sites, landing pages and ad content from one prompt.

LLMsBackendIntegrations
wandit.dev ↗

04 GitHub

Built in the open.
Commit by commit.

0 contributions in the last year

DennixAI ↗

05 Skills

A mixture
of experts.

Like the MoE layers I build, every problem gets routed to the experts it needs. Pick a task, or watch the router choose.

Input token

Ship a RAG chatbot
Router top-2
Shared expertalways on

Python · Git · Linux · NumPy · pandas · reading papers

Research

Prototyping, evaluation, benchmarking · Transformers, MoE, reasoning

Training & inference

PyTorch · JAX · DDP · AMP · gradient checkpointing · MLOps

Applied AI

LangChain agents · RAG · conversational memory · AI integration

Backend

Node.js · FastAPI · REST · PostgreSQL · Docker · CI/CD

Web & mobile

TypeScript · React Native · front-end / back-end

ML domains

RL · NLP · computer vision · generative AI · data science

GPU & systems

CUDA · C/C++ · Rust · RunPod · Lambda · Vast.ai · Verda

Hardware

Arduino · ESP32 · CAD & 3D printing · battery management systems

06 Education

CESI — Nanterre

Third year of computer science · in progress

Baccalauréat, Mathematics

Mention Très Bien — highest honours

07 Languages

Frenchnative
Englishfluent

08 Currently obsessed with

AI researchReinforcement learningInference engineeringGPU programmingDistributed computingHardwareAI startups

09 Contact

Let's build
the next layer.

I'm looking for an 8–16 week internship in AI, machine learning or data science, starting December 18, 2026. Paris or remote.

ryad.dn@proton.me ↗
Portrait of Ryad Denni
Ryad Denni21 · Paris · AI / Backend