English · Français · العربية

I'm Ismail, an AI engineer and researcher based in Paris, France 🇫🇷 (originally from Fès, Morocco 🇲🇦)

I learn by building. I reimplement papers from scratch (vision transformers, GPTs, an autograd engine, voice-conversion models) so I understand how they work instead of just calling an API. I turn that understanding into practical AI systems: document extraction, agent evaluation, real-time audio, and independent products such as Qlaver, Kinetra and Glitch.

• Graduate of École Centrale de Lyon (Master’s in Artificial Intelligence) and ENSA Fès (Computer Science Engineering degree, top of the class)
• Focused on agentic AI, multilingual RAG and LLM evaluation in production
• Exploring multimodal learning (vision + language) and real-time generative audio
• Motto: build it from scratch first, reach for the library second.

Neural Lab deep learning from first principles	Agentic AI LLMs wired into real workflows	Product Engineering mobile apps, SaaS & software that ships
Transformers, ViTs & GPTs rebuilt from papers Multimodal vision + language (captioning, cross-attention) Generative audio: real-time voice conversion Graph NNs (GCN, GAT, MPNN) Autograd & optimizers by hand	LLM apps & multi-agent systems RAG: hybrid search, re-ranking, vector DBs Agent harnesses: tools, MCP, Claude Code skills Evaluation & feedback loops for agents Fine-tuning (LoRA) & prompt engineering	SaaS backends: FastAPI, PostgreSQL, async Web apps: React, Next.js, TypeScript Mobile apps: Android with Kotlin Cloud & DevOps: AWS, Azure, Terraform, CI/CD Tooling in Go & Python
PyTorch · Lightning · Accelerate · W&B	LlamaIndex · LangChain · Hugging Face · Qdrant	FastAPI · React · Kotlin · Docker · AWS

_{Explore my tech stack in 3D → click a technology to see the projects and jobs I used it in.}

Languages

AI / Machine Learning

Vector DBs, MLOps & Experiment Tracking

Backend, Cloud & DevOps

Frontend & Mobile

Products I build in my spare time, to take AI from a notebook to something people can use. These three are the ones I keep working on to make them ready for real users. They are private for now: I see real potential in them, so their code stays closed.

Project	Description	Stack
MusicGen	Controlled comparison of Transformer, LSTM and RNN models for symbolic music generation.	PyTorch · Keras · Streamlit
🔧 MyTorcha	Minimal deep learning framework with DAG-based autograd, written in pure Python.	Python · NumPy
📐 cadquery-agent	Local text/image agent that runs on open-weight models and generates validated CadQuery CAD code.	Python · LLMs
📐 tiny-img2cad	Image-to-CadQuery model: a ResNet18 encoder and a Transformer decoder that write the code for a CAD part from its picture.	PyTorch
👁️ vit-image-classifier	Vision Transformer written from scratch: attention heatmaps, Streamlit demo, 86 tests, 97% on shapes and 86% on PlantVillage.	PyTorch · Streamlit
🎯 yolo-object-detector	YOLOv1 from the paper: network, loss, NMS and mAP, trained on synthetic shapes, MNIST, PASCAL VOC and Mapillary traffic signs.	PyTorch
🧩 segnet-semantic-segmentation	SegNet semantic segmentation on toy shapes, Oxford-IIIT Pet and Mapillary Vistas, with tests and a Kaggle/Colab notebook.	PyTorch
🖼️ text2image	Twelve text-to-image architectures (SD U-Net, DiT, MMDiT, FLUX, SANA, MaskGIT, VAR, ...) trained and benchmarked side by side.	PyTorch
🎨 imagegen	Image generation with GANs (DCGAN, SNGAN) and diffusion (DDPM/DDIM U-Net) on MNIST, CIFAR-10 and AFHQ.	PyTorch
🎙️ stt-workbench	Speech-to-text workbench: CTC, LAS, hybrid CTC/attention, RNN-Transducer and wav2vec2/Whisper fine-tuning, with WER/CER evaluation.	PyTorch
🔊 tts-workbench	Text-to-speech workbench: Tacotron 2, Transformer-TTS, FastSpeech 2, VITS and a HiFi-GAN vocoder, with MCD/F0/MOS evaluation.	PyTorch
🧑‍💼 hr-jev	One GPT-2, written from scratch, predicts four HR outcomes (attrition, job level, satisfaction, income) from a plain-words employee description.	PyTorch · Gradio
💬 text-classification	Emotion classification with RNN, LSTM and GRU built with nn.* and from scratch; 12 models at 91-92% test accuracy.	PyTorch
🕹️ cartpole-rl · lunarlander-rl · pong-rl	Every RL family (tabular, DQN variants, REINFORCE, A2C, PPO, SAC, pixel CNN policies) compared on one environment each, with before/after GIFs and curves.	PyTorch · Gymnasium
🏔️ mountaincar-rl · pendulum-rl · frozenlake-rl	Sparse-reward, continuous-control and classical tabular RL (value/policy iteration, SARSA, Q-learning, Dyna-Q, DDPG/TD3/SAC) with heatmaps and exact-optimum checks.	PyTorch · Gymnasium
🧪 gan-workbench	Ten GAN variants (GAN to WGAN-GP, SNGAN, pix2pix, CycleGAN) on one training loop with FID/IS/IoU evaluation and training-progress GIFs.	PyTorch
🌫️ diffusion-workbench	Diffusion and flow models (DDPM, DDIM, v/x0-prediction, score SDE, flow matching, EDM) with U-Net and DiT backbones, FID vs NFE.	PyTorch
🩻 unet-image-segmentation	U-Net with every architectural switch made explicit (skips, residual blocks, strided vs pooled), six variants scored with IoU/Dice on shapes and Oxford-IIIT Pet.	PyTorch
😮 speech-emotion-recognition	Eight audio classifiers (WavLM, CRNN, CNNs, MLP, SVM) on RAVDESS, evaluated on unseen speakers; best 82.5% accuracy.	PyTorch · Hugging Face
🔤 en_2_ar_translator	English-to-Arabic Transformer machine translation with shared BPE, beam search and BLEU/chrF evaluation.	PyTorch
📨 naive_little_rag	A simple, readable RAG system that shows each step of retrieval-augmented generation.	Python · Vector search

➡️ See all repositories

Here's a fun exchange I had with Tiny LLM (formerly MT-GPT), the GPT I built from scratch:

• Top Graduate: 2024 Computer Science Engineering class, ENSA Fès
• 1st in Morocco · 121st worldwide: IEEEXtreme global coding competition
• 2nd place: IEEE R8 Entrepreneurship Initiative, with Equinox, a fitness app concept

I'm always happy to talk about deep learning, GenAI, research collaborations, or startup ideas.

All my links in one place: benisalla.github.io/my-link-in-bio