Srijani Das

• Conducting research and developing CNN-based models for spatially resolved wind pressure prediction on building surfaces, trained on experimental wind tunnel data from the TPU Aerodynamic Database.

• Developed a full-stack embeddable web widget with a configurable JSON-driven architecture, supporting multiple deployment modes and dynamic UI state management.
• Built the interruption-handling pipeline for a voice AI telephony system implementing real-time voice activity detection for barge-in, pre-speech audio buffering.
• Implemented single-retry fallback logic for truncated transcripts along with WebSocket audio streaming.
• Evaluated speech-to-text and text-to-speech approaches for a voice AI telephony application through comparative testing and proof-of-concept experiments.

Built a Node.js/Express tool using a headless Chrome instance (Puppeteer) to export ChatGPT and Claude conversations to Markdown, applying platform-specific extraction strategies: parsing embedded JSON for ChatGPT and intercepting runtime API responses for Claude.

Diagnosed and fixed an intermittent race condition in async response interception by restructuring page.waitForResponse() to resolve before navigation could complete.

Built an ML classifier for stars, galaxies, and quasars using SDSS photometric data.

Designed a custom three-stage hybrid model (PRISM) that combines a color graph encoder, mutual information attention, and calibrated LightGBM, achieving 98.04% test accuracy — outperforming Random Forest (97.82%) and XGBoost (97.85%) baselines.

Developed an interactive dashboard for real-time predictions on user-input photometric data.

Developed a full-stack recruitment consulting portal for a startup with secure authentication and a responsive admin panel for job and candidate management.

Built RESTful API backend using Node.js with SQLite database for CRUD operations on job postings.

Built a local Python application that transcribes audio files and generates AI-powered summaries at multiple detail levels (short, medium, detailed). Implemented privacy-focused architecture with all processing performed locally using Whisper for speech-to-text and BART for summarization.

Developed user-friendly Streamlit interface with downloadable transcript and summary export functionality. Supports multiple audio formats (MP3, WAV, M4A, FLAC, OGG).

Conducted research on automated plant disease detection using deep learning under faculty supervision. Implemented and compared CNN, Vision Transformer (ViT), and hybrid ViT+CNN architectures for multiclass classification on the PlantVillage dataset, achieving ~99% accuracy.

Applied Grad-CAM visualizations to interpret model predictions and enhance explainability. Research was extended into a published paper in Frontiers in Artificial Intelligence (March 2026).