knowledge/ml_projects.md

Machine Learning Projects & Experience

DevelopersHub Corporation — Advanced AI/ML Engineering Internship (Mar–Apr 2026, remote)

Arman was selected for the Advanced track and delivered three production-grade projects:

1. Fine-tuned BERT-base-uncased on the AG News dataset (120K samples) → 94.74% accuracy and 94.75 macro F1 (batch size 16, learning rate 2e-5, fp16 mixed precision, trained on an RTX 4060 with the HuggingFace Trainer).
2. End-to-end Customer Churn Prediction pipeline on IBM Telco (~7K rows): ColumnTransformer preprocessing plus GridSearchCV (5-fold, F1) over Logistic Regression and Random Forest, with ROC and confusion-matrix evaluation.
3. LLM support-ticket auto-tagger: benchmarked zero-shot, few-shot, and fine-tuned approaches — the fine-tuned model reached 99.82% accuracy.

Independent research — IoT Intrusion Detection (2024–2025)

Arman authored a research paper, "Optimized Deep Learning Intrusion Detection for Enhanced IoT Network Security": a hybrid CNN-LSTM intrusion detection system with NSGA-II Jumping-Gene multi-objective feature selection on the CICIDS-2017 dataset → 99.03% accuracy with roughly 5× faster training than MLP/CNN baselines. He implemented CNN, RNN/LSTM, and BERT models in PyTorch with full preprocessing, Random Forest feature ranking, and per-class precision/recall/F1 analysis.

Live demo on HuggingFace

The BERT AG-News classifier is deployed as a live Gradio Space: huggingface.co/spaces/ArmanXAI/ag-news-classifier-demo (dark-themed UI with confidence bars). The model repo lives on Arman's HuggingFace profile: huggingface.co/ArmanXAI.

Other projects

• EmotionCNN — a Happy/Sad facial expression classifier built from scratch in PyTorch (64×64 grayscale images), with a CPU vs GPU training/inference benchmark on an RTX 4060.
• General Health Query Chatbot — LLaMA 3.1 8B via Groq + HuggingFace.
• Smart Attendance System — IoT + biometric attendance on ESP32-CAM (team of 3; Arman owned the audio module).

Freelance work (Upwork, 2024–present)

B2B lead generation, web scraping, and structured-data projects for international SMB clients, with consistently positive feedback. Built reusable Python scrapers and AI-assisted enrichment workflows (ChatGPT, Notion, Make/Zapier).