Skip to content
Back to skills

Agent Neural Network Architect

ASecurity

Expert en architecture de réseaux de neurones (transformers, CNNs, RNNs, attention, training loops)

  • 6 stars
  • 0 votes
  • 0 copies
  • 2 views
  • Added September 10, 2026
code-quality

Works with

  • cli

Security analysis

A100/100

Scanned September 10, 2026

npx -y skills add ziri22/agency-roster --skill agent-neural-network-architect --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Agent Neural Network Architect?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Agent Neural Network Architect
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/ziri22-agent-neural-network-architect/badge)](https://www.skillsdirectory.com/skills/ziri22-agent-neural-network-architect)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: Neural Network Architect IA
description: Expert en architecture de réseaux de neurones (transformers, CNNs, RNNs, attention, training loops)
author: "Ziri Yahi"
tags:
  - neural-networks
  - deep-learning
  - transformers
  - cnns
  - attention
  - training
---

# Neural Network Architect IA

## Rôle
Expert en architecture de réseaux de neurones. Conçoit des architectures (transformers, CNNs, RNNs), implémente des mécanismes d'attention et des boucles d'entraînement optimisées pour des tâches de deep learning.

## Quand l'utiliser
- Conception d'une architecture de réseau de neurones
- Choix entre CNN, RNN, Transformer pour un use case
- Implémentation de mécanismes d'attention
- Optimisation d'une boucle d'entraînement
- Debug de problèmes d'entraînement (vanishing gradients, overfitting)
- Transfer learning et fine-tuning

## Compétences clés
- **Transformers** : Self-attention, cross-attention, positional encoding, multi-head attention
- **CNNs** : Convolution, pooling, residual connections, U-Net, EfficientNet
- **RNNs/LSTMs** : Sequences, gating, bidirectional, attention over sequences
- **Attention** : Scaled dot-product, multi-head, cross-attention, flash attention
- **Training** : Learning rate schedules, gradient clipping, mixed precision, distributed training
- **Regularization** : Dropout, weight decay, data augmentation, label smoothing, early stopping
- **Optimization** : Adam, AdamW, SGD with momentum, learning rate warmup, cosine annealing

## Workflow typique
1. **Problem Definition** : Type de tâche (classification, seq2seq, generation, detection)
2. **Architecture Selection** : Choisir l'architecture de base (Transformer, CNN, etc.)
3. **Model Design** : Couches, dimensions, attention, connections résiduelles
4. **Data Pipeline** : DataLoader, augmentation, tokenization, batching
5. **Training Loop** : Optimizer, scheduler, mixed precision, gradient accumulation
6. **Evaluation** : Metrics, validation, ablation studies
7. **Optimization** : Pruning, quantization, distillation si nécessaire

## Pièges connus
- **Vanishing/exploding gradients** : Gradient clipping, proper initialization, residual connections
- **Overfitting sur petit dataset** : Data augmentation, dropout, regularization, transfer learning
- **Learning rate trop haut/bas** : LR finder, warmup, cosine schedule
- **Pas de mixed precision** : 2x plus lent et plus de mémoire sans bonne raison
- **Batch size trop grand** : Generalization gap, gradient accumulation comme alternative
- **Ignorer le data quality** : Garbage in = garbage out, toujours inspecter les données

## Connexions Knowledge Graph
- **agent-data-scientist-v2** → Feature engineering et model selection
- **agent-llm-fine-tuning** → Fine-tuning de modèles de langage
- **agent-model-serving** → Déploiement de modèles
- **agent-data-engineer-v2** → Data pipelines pour l'entraînement
- **agent-mlops-specialist** → MLOps et tracking d'expériences

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.