I'm Alif Ilham Madani

Accelerated the PyTorch execution graph utilizing torch.compile and engineered a custom Triton INT4 GEMV kernel. Synergistically integrated flex_attention and Triton autotuning to drive significant improvements in Time-To-First-Token (TTFT) and decode throughput.

Engineered an end-to-end TinyML audio classification pipeline deployed onto an Arduino Nano 33 BLE. Implemented Quantization-Aware Training (QAT) and structured channel pruning to achieve an 18% physical inference speedup directly on MCU silicon.

Engineered a 2D SLAM system from scratch fusing 4-channel wheel encoder telemetry with Hokuyo Lidar. Architected a Sequential Monte Carlo algorithm with dynamic roulette-wheel resampling to achieve large-scale loop closure.

Constructed theoretical and empirical Roofline models to analyze operational intensity and peak TFLOPS across 10 deep neural network architectures, isolating performance bottlenecks within heterogeneous CPU and GPU environments.

Engineered a full-stack TypeScript web application integrating Vision-Language Models to perform zero-shot visual analysis of biomechanical form in workout imagery, utilizing an automated prompt-engineering pipeline.

Architected a 1D Convolutional Triplet Network optimized via Triplet Margin Loss to map complex improvised audio sequences into a robust semantic embedding space for k-NN classification.

Engineered an inertial measurement unit (IMU) time-series classification system to recognize complex 6-DOF spatial gestures using 3-axis accelerometer and gyroscope data. Architected K-Means clustering for vector quantization and trained left-to-right HMMs with Baum-Welch algorithm, implementing all forward-backward calculations in log-space to eliminate numerical underflow across long temporal sequences.

Engineered a vision-based obstacle detection pipeline utilizing YCrCb color space to decouple luminance from chrominance across varying lighting conditions. Trained multivariate Gaussian distribution models via MLE with quantized 6-bit Look-Up Tables for real-time inference, and implemented 4th-degree polynomial regression achieving 12.78 cm RMSE on distance estimation.

Engineered an end-to-end anomaly detection pipeline classifying cardiovascular heartbeats from low-fidelity 128x128 grayscale images, overcoming extreme class imbalance (<1% minorities) and severe domain distribution shifts. Architected custom signal-reconstruction module translating 2D rasterized images to 1D time-series, and trained deep 1D ResNet-34 with Focal Loss and stratified k-fold cross-validation, boosting performance from 0.43 baseline to robust 0.836 test score.

Architected a 5-stage NLP optimization pipeline combining 300-dimensional FastText embeddings and XGBoost gradient boosting for 5-class sentiment classification, achieving 93.5% test accuracy. Engineered supervised pseudo-labeling augmentation to leverage sparse datasets where 58% of 59,000+ training samples were unlabeled, outperforming traditional KNN Label Propagation while executing rigorous ablation studies.