Elnur Mammadli

SWEEP-Net: domain-agnostic EEG affective-state decoding

Lead Researcher · Baku · ongoing

Architecting a low-power, neuromorphic pipeline for cross-subject 5-class EEG emotion recognition — integrating decoupled contrastive representation learning, Multiple Instance Learning (MIL), and Spiking Neural Network (SNN) distillation.

• Spatiotemporal features. A 5D Atrous-MobileNet encoder over continuous wavelet transform (CWT) maps, with zero-FLOP bi-directional Temporal Shift Modules (TSM) for temporal context; trained with Decoupled Contrastive Learning (DCL) plus a momentum queue to prevent positive-pair preemption and recover a continuous affective geometry.
• Biometric debiasing & subject cleansing. Euclidean Alignment (EA) sequence whitening to suppress scalp-impedance artifacts — cutting proxy subject-identification leakage from 87.10% to 64.10% — with a covariance-rank spectral gating test that conditionally nullspace-projects out linearly separable subject fingerprints.
• Temporal routing (MIL). A tri-branch SwiGLU-MIL aggregation head — Average, LogSumExp, and Attention branches — to resolve temporal label pollution, with focal loss and per-class target-entropy regularisation isolating brief affective climaxes from long neutral baselines.
• Neuromorphic deployment. A hybrid spatio-temporal knowledge-distillation (ST-KD) framework that distils continuous activations into an Adaptive Leaky Integrate-and-Fire (ALIF) spiking network, tuning membrane-leak decay and threshold dynamics to maximise Synaptic-Operation (SynOps) efficiency for FPGA deployment.

TrashAI — real-time waste classification API

Python · YOLOv6 · Flask · Redis · 2021–2024 · GitHub

A multilingual Flask web app and REST API serving real-time waste classification, backed by a custom YOLOv6-S detector trained on a curated 14K-image dataset — 82.4% AP on Plastic and 80.1% AP on Paper, beating YOLOv4 baselines — with JWT authentication and Redis-backed rate limiting on the inference endpoint.

Deep learning framework from scratch

Python · NumPy · 2020 · GitHub

CNN/MLP architectures with hand-written forward/backward passes and optimisers in raw NumPy — no autograd, no framework. Proof that I understand the math, not just the API.

Handwritten OCR & text segmentation (NTO Finals)

Detectron2 · PyTorch · CRNN · 2022 · GitHub

Mask R-CNN text-line polygons → CRNN (ConvNeXt/ResNet-34 + BiLSTM, CTC loss). Reached a modified CER of 0.222 — 16th of 42 national finalist teams.

2GIS Road Sign Labeling (AIIJC)

Python · EfficientNetB2 · Deep Learning · Nov 2021 · GitHub

Constructed a multiclass-multilabel image classification model using EfficientNetB2 to categorize road signs across 6 basic headings. Applied a comprehensive data augmentation pipeline (optical distortion, color jitter, blur) to significantly enhance model generalization and robustness.

Silicon Spatula

Java · MVC Architecture · May 2026 · GitHub

Programmed a comprehensive restaurant management simulation driven by a custom Java Swing GUI. Enforced strict MVC architecture utilizing custom generic queues and domain-aware exception hierarchies for robust state management and thread safety.

Smart House IoT Controller

C# · Xamarin · Raspberry Pi · 2019 · GitHub

Built a native Android application via Xamarin to remotely control GPIO pin states on a Raspberry Pi 3B+ server via asynchronous HTTP endpoints.