Riddhi Sharma

B.Tech graduate in Artificial Intelligence with hands-on industry experience delivering production-grade AI systems, including retrieval-augmented generation (RAG) pipelines and real-time audio deep learning models, at Samsung R&D Institute India, Delhi. Proficient in Python, Machine Learning, Deep Learning, Natural Language Processing, and end-to-end model deployment, with a track record of quantifiable performance improvements in enterprise-scale AI infrastructure.

Intern - Samsung R&D Institute India - Delhi

(2026-01 - 2026-07)

Project 1: Codebase Indexing for Internal AI Assistant (Cline SR) Project 2: Real-Time Audio Classification Pipeline (Music Detection)

• Architected and optimized a repository-level RAG pipeline for Cline SR, Samsung’s internal AI coding assistant, enabling semantic context-aware code search across large Android repositories containing 5,000+ files.
• Engineered a codebase indexing workflow utilizing vector embeddings, systematic tokenization, and hierarchical file parsing to enable accurate semantic query mapping across multi-module Android projects.
• Diagnosed and resolved a critical file-reference consistency bug where the AI resolved different files across repeated queries on the same codebase; isolated deterministic multi-query variance, reducing resolution errors by ∼40% and documented mitigation frameworks for enterprise deployment.
• Gained deep hands-on exposure to LLM-based developer tooling, retrieval-augmented generation pipelines, and production AI assistant infrastructure at enterprise scale.
• Redesigned the end-to-end audio ML pipeline from a file-based batch workflow to a real-time streaming architecture, eliminating the need to download WAV files prior to inference and significantly reducing end-to-end latency.
• Implemented live audio capture directly from the speaker/microphone input, replacing the previous offline WAV download → spectrogram → model inference flow with an always-on streaming pipeline.
• Applied RMS normalization (target: 0.1 RMS) to incoming audio streams to ensure consistent input amplitude, improving model robustness across varying audio conditions.
• Engineered a real-time spectrogram generation stage using Librosa that converts live normalized audio frames into high-resolution spectrograms on-the-fly before feeding them into the deep learning inference model.
• Built a multi-class audio classifier targeting four categories: Music, Speech, Music + Speech (overlap), and Noise, enabling fine-grained content detection beyond binary music/non-music classification.
• Processed a dataset of 10,000+ audio samples for downstream model training, porting the pipeline to a modern spectrogram-based architecture aligned with production-grade deep learning standards.

B.Tech - Artificial Intelligence - Banasthali Vidyapith (2026-05)

Class XII - Sir Padampat Singhaniya School (2022-04)

Class X - Sir Padampat Singhaniya School (2020-04)