Production Causal Recommendation Engine & Lakehouse (21M+ records, SASRec Transformers, LightGCN, PySpark Medallion, pgvector HNSW, ONNX Runtime)

Deep learning system for football match outcome, exact score and player performance prediction using LSTM, MLP, Elo ratings, Poisson models and Monte Carlo simulation.

AI crop disease detector for farmers: photo diagnosis with Bangla/English advice (MobileNetV2 + FastAPI + Gemini)

Claude skills for ML: turn research papers into clear breakdowns and minimal, runnable PyTorch code

An AI-powered web application that detects AI-generated images using deep learning, PyTorch, and computer vision. Built with PHP, Python, MySQL, and Docker, featuring image analysis, confidence scores, and an intuitive web interface.

RSNA Knee Abnormality Detection (Kaggle 2026): CoAtNet MIL fine-tuned with OAI MOAKS expert readings. Private LB 0.943, rank 345/6,425 (top 5.4%).

PyTorch classifier for Chickenpox, Measles, Monkeypox and Normal skin images, with duplicate-safe data splitting and a minimal Tkinter app. Educational project, not a medical device.

Small Audio AI project for environmental sound classification using the ESC-50 dataset and a pretrained Audio Spectrogram Transformer (AST) model.

A 2-layer attention-only transformer trained from scratch in PyTorch, reproducing the induction-head phase transition, with TransformerLens attention analysis.

End-to-end Computer Vision pipeline for employee activity recognition and productivity analysis using YOLO11, BoT-SORT, SlowFast, MediaPipe, and multi-model fusion.

A Llama-style GPT built from scratch in PyTorch: byte-level BPE tokenizer, RoPE, RMSNorm, SwiGLU, KV-cache sampling.

Domain-specific restoration pipeline for broadcast anime video on the Re-Anime600 benchmark (AINIME Challenge @ WACV 2027 Workshop).

Federated learning with noisy labels through label correction, distillation and noise-aware aggregation.

RSNA Knee Abnormality Detection: Commercial Clinical MRI Diagnostic Engine & Multi-Modal Semi-Supervised Deep Learning Pipeline

Pose-based human fall detection from 360° video using YOLO pose estimation and temporal deep-learning models.

一套面向实践的 **OpenAI Agents SDK（Python）** 中文教程。从 `Agent` + `Runner` 两个概念起步，一路讲到工具、会话、流式、多智能体编排、人工介入、安全防护、追踪与测试。

Knee osteoarthritis grading on the Kellgren–Lawrence (KL) scale from X-rays, with custom CNNs trained from scratch, Grad-CAM, and an in-browser app.

Explainable ensemble deep learning framework for automated sugarcane leaf disease classification using attention mechanisms, weighted soft voting, Grad-CAM, and a Streamlit prototype.

Twelve to One: lead-robust deep learning for multi-label ECG diagnosis and wearable triage on PTB-XL (NTU BS6204)

Medical image classification · 97% acc on NIH Chest X-Ray 14 · PyTorch · Focal Loss · Grad-CAM

A Universal Transformer that solves 4x4 Sudoku by repeatedly applying one shared block and learning when to stop

GPU-accelerated data-center power forecasting (LSTM/TCN with CUDA mixed precision, quantile P10/P50/P90) with rolling backtests, CPU vs GPU benchmarks, experiment tracking, FastAPI serving and an LP carbon-aware compute scheduler.

See what a deployment compiler did to your model: source and compiled graphs side by side, every fused kernel traced to its source ops. OpenVINO and TensorRT, runs in the browser, uploads nothing.

Computer Vision pipeline for forensic tracking and Loss Prevention using YOLOv8, ByteTrack, and Person ReID.

PyTorch Fundamentals — solutions to all 10 exercises + challenge (device-agnostic: CUDA / MPS / CPU)

Human action recognition using PyTorch and MediaPipe, featuring CNN-LSTM and pose-based models, ablation studies, and reproducible experiments.

Educational project exploring pediatric pneumonia classification from chest X-rays with a calibrated, robust, and explainable CNN.

Benchmarking backdoor data-poisoning attacks (5 datasets × 5 architectures × 3 poison rates) and embedding-based poison detection — minor project on data poisoning using self-supervised representation learning.

Agentic AI system for brain MRI analysis: segmentation, uncertainty estimation, vision-language report generation, and automatic human-review flagging.

A synthesizable FlashAttention engine in SystemVerilog: verified bit-exact against a golden model, compiled to WebAssembly, running live in your browser.