class Mahdi:
role = "ML Engineer / Researcher / Systems Tinkerer"
focus = ["Privacy-Preserving ML", "Federated Learning",
"Semi-Supervised Learning", "Distributed Systems"]
exploring = ["Mechanistic Interpretability", "Differential Privacy",
"Adversarial Robustness", "Confidential Computing"]
motto = "Models should learn from data, not memorize it."Machine Learning & Deep Learning
Architectures : CNN, ResNet, U-Net, Transformer, ViT, GAN, VAE, Diffusion, GNN
Paradigms : Supervised, Self-Supervised, Semi-Supervised, Few-Shot, Meta
NLP : BERT, GPT-family, T5, LLaMA, RAG, fine-tuning, LoRA, PEFT
CV : Detection, Segmentation, OCR, Pose, Multi-modal (CLIP, BLIP)
RL : PPO, DQN, SAC, MARL, RLHF
Optimization : Adam, AdamW, SGD-momentum, LR schedules, mixed precision
Privacy-Preserving ML & Security
-792EE5?style=for-the-badge&logo=pytorch&logoColor=white){kind=icon}
Privacy : Differential Privacy (DP-SGD, PATE), k-Anonymity, l-Diversity
Federated : FedAvg, FedProx, FedSGD, Secure Aggregation, Personalization
Cryptography : Homomorphic Encryption (CKKS, BFV), MPC, ZK proofs (basics)
Threat Modeling : Membership Inference, Model Inversion, Data Poisoning
Adversarial ML : FGSM, PGD, AutoAttack, Certified Defenses, Adv. Training
Confidential Comp. : Intel SGX, AMD SEV, AWS Nitro Enclaves
Auditing : LiRA, Privacy Accounting (RDP, GDP), Canary Insertion
Distributed Systems & Infrastructure
Concepts : CAP, PACELC, Consensus (Raft, Paxos), CRDTs, Vector Clocks,
Eventual Consistency, Sharding, Replication, Service Mesh,
Circuit Breakers, Backpressure, Idempotency, Sagas
Patterns : Event Sourcing, CQRS, Pub/Sub, Leader Election, Bulkheads
Storage : etcd, ZooKeeper, Cassandra, ScyllaDB, MinIO, Ceph
Languages, Backend, Cloud (click to expand)
Languages
Backend
Cloud / DevOps
$ ./loss_curves --model=research_progress
loss
│
1.0 ●╲
│ ╲
0.8 │ ╲●─╲
│ ╲●─╮
0.6 │ ╲●──●─╮
│ ╲●──●─╮
0.4 │ ╲●──●──●─╮
│ ╲●──●──●──●
0.2 │
│ papers read · models trained · privacy preserved
0.0 └────────────────────────────────────────────────────▶ epoch
0 100 200 300 400 500 600 700 800
@@ ongoing experiments @@
+ DP-SGD with adaptive clipping — does ε stay sane at scale?
+ Federated SSL on non-IID data — pseudo-label drift across clients
+ Membership inference vs label-DP — sharper bounds, better defenses
- "just throw more compute at it" — usually wrong, occasionally right
! Reading: anything by Dwork, McMahan, Goodfellow, Olah
