Harsh Soni

Computer Science undergraduate (GPA 8.61/10) and current research intern at DRDO, specialising in machine learning, deep learning, and generative AI. Builds end-to-end ML systems — spanning signal-domain pipelines, embedding-based retrieval, and LLM-powered applications — with hands-on experience in PyTorch, classical ML, transfer learning, fine-tuning, and retrieval-augmented generation. Strong foundation in feature engineering, model training, and quantitative benchmarking under realistic, noisy conditions.

Research Intern - CASDIC, DRDO — Defence Research & Development Organisation - Bengaluru, India

(2026-05 - 2026-07)

Radar Pulse Deinterleaving & Emitter Classification (ELINT) using Python, NumPy, scikit-learn, DBSCAN, and Random Forest

• Built a multi-stage ELINT pipeline (signal simulator → DBSCAN deinterleaver → feature extractor → emitter-type classifier) on a physics-based pulse simulator generating ∼189K pulse-descriptor words across 25 emitters and 5 radar types
• Implemented density-based deinterleaving on a weighted (AOA, FC, PW) feature space, achieving 94.3% pulse-assignment accuracy and recovering 22/25 emitters under clean conditions
• Engineered 24 per-cluster statistical and behavioural features (PRI mean/CV/span, modulation type, frequency-agility flag) with strict feature selection that excludes position-revealing inputs to prevent data leakage
• Benchmarked Logistic Regression, KNN, SVM, Random Forest, and XGBoost across four noise tiers, selecting Random Forest as the final classifier (95% type-classification accuracy on clean data; highest mean accuracy across all tiers), with quantified degradation curves for honest robustness reporting
• Authored an error-attribution stage decomposing mistakes into deinterleaving vs. classification error to surface intrinsically separable vs. ambiguous radar pairs

B.Tech - Computer Science and Engineering - Kalinga Institute of Industrial Technology (KIIT) (2023-09 - 2027-05)

Senior Secondary - CBSE - Kendriya Vidyalaya, ASC Centre (2019-04 - 2022-03)