TUSHAR SINGLA

Real-time sign language translation powered by temporal deep learning & MediaPipe vision.

End-to-end accessibility platform solving real-world communication barriers for the deaf and mute community. Engineered a custom deep learning sequence model in PyTorch paired with a 21-point MediaPipe hand-landmark extraction pipeline. Overcame complex multi-frame gesture dynamics with stacked LSTM temporal modeling, deployed with real-time speech synthesis, live captioning, and full-stack web integration.

On-device Vision Transformer OCR & local quantized model workflows.

Awarded Top 5 at Eclipse Hackathon (Thapar University). An offline-first translation and micro-learning engine integrating local Manga-OCR (Vision Transformer) and quantized models on-device via Ollama. Eliminates cloud API dependencies with high-performance local inference, canvas text inpainting, and IndexedDB caching for seamless offline workflows.

Automated contract intelligence & NLP risk classification.

Awarded 2nd Place at Brainwave 6.0 (Delhi Technological University). An automated NLP intelligence platform that ingests complex legal agreements, contracts, and NDAs to extract and classify high-risk liability, indemnity, and termination clauses in seconds. Demonstrates applied AI integrated directly into legal document workflows.

High-throughput vector retrieval & hybrid search without framework overhead.

Engineered an end-to-end retrieval-augmented generation pipeline using foundational vector similarity and cosine distance mathematics. Implemented zero-overhead direct SDK integrations, custom semantic chunking, and low-latency embeddings for enterprise knowledge workflows.