Disha Saini

About

Third-year AI/ML undergraduate researcher focused on deep learning, computer vision, audio machine learning, multimodal learning, and probabilistic modelling. Experience includes designing and evaluating neural architectures, running controlled experiments and ablations, and building research prototypes for audio source separation, crowd classification, and motorsport telemetry analysis.

Experience

• Undergraduate Minor Research Project
  Freelance

  Graph-Semantic-Net: Text-Guided Audio Source Separation. Formulated and evaluated a hybrid Harmonic GCN–U-Net architecture conditioned with CLAP-based cross-attention for text-guided audio source separation and semantic source steering. Implemented the end-to-end signal processing and modelling pipeline, incorporating STFT/iSTFT representations, graph-structured harmonic representations, and cross-attention multimodal conditioning. Evaluated generalization to unseen audio classes and applied structured parameter pruning, obtaining approximately 31% parameter reduction relative to the baseline while assessing inference efficiency tradeoffs.

• Co-authored Research Paper
  Freelance

  Real-Time Crowd Classification via Deep CNNs. Designed and evaluated a compact three-layer Convolutional Neural Network for real-time binary crowd versus non-crowd classification. Executed controlled comparative experiments assessing activation functions and data augmentation strategies, accompanied by systematic ablation and empirical evaluation. Reported classification performance including approximately 82% accuracy/F1-level score, an AUC of 0.88, and 87% recall on the crowded class.

• AI/ML Pipeline Engineer
  Motorsport Analytics Startup

  Developed and maintained data-driven telemetry pipelines focused on Formula One tyre degradation dynamics and predictive performance modelling.

Education

• B.Tech in Computer Science and Engineering
  NIELIT Deemed to be University
  Artificial Intelligence & Machine Learning · 2024 – 2028

Skills

• C <3 months
• Computer Vision <3 months
• Deep Learning <3 months
• Python <3 months
• NumPy <3 months
• Pandas <3 months
• Streamlit <3 months
• scikit-learn <3 months
• Flask <3 months
• PyTorch <3 months
• TensorFlow <3 months
• Scipy <3 months
• CNN (Convolutional Neural Network) <3 months
• Plotly <3 months
• Reinforcement Learning <3 months
• RAG <3 months
• LLMs <3 months
• Transformers <3 months
• XGBoost <3 months
• Audio Signal Processing <3 months
• Multimodal Learning <3 months
• CatBoost <3 months
• Model Explainability <3 months
• probabilistic modeling <3 months

Tools / apps / platforms

• Docker <3 months
• Git <3 months
• Jupyter Notebook <3 months
• MySQL <3 months

Projects

• VAKYAI: AI Communication Intelligence Platform
  Multimodal Machine Learning

  Formulated a multimodal assessment framework integrating acoustic characteristics, lexical patterns, and dialogue structure to extract structured communication-performance metrics.

• JEEVAN: AI Emergency Response System
  Machine Learning

  Designed an end-to-end machine learning pipeline for automated accident detection, scene event classification, and real-time coordination across emergency services.

• Object Detection and Classification Framework for Indian Road Video
  YOLOv8, Indian Driving Dataset (IDD)

  Developed an object detection and classification framework using YOLOv8 trained on the Indian Driving Dataset across 35 traffic classes under unstructured road conditions.

🏆 Achievements & awards

• Top 10 Finalist, TrackShift Formula One Hackathon · 2026
• Top 3, Eureka! College Startup Competition · 2026
• Co-author & Presenter, 4th NIELIT International Conference (NICEDT–2026) · 2026

  Presented crowd classification research.

⚽ Extracurricular activities

• Vice President, Entrepreneurship Cell
• Coordinator, Vaktavya (Public Speaking and Oratory Club)

  Leading Vaktavya Reach, an initiative introducing structured communication workshops to secondary school students.