Projects
Selected AI projects covering retrieval systems, model evaluation, and sequence generation. Built around practical ML engineering and deployment.
Audio2Text RAG System
PROJECT_2025
An end-to-end Retrieval-Augmented Generation service that ingests documents and audio, normalizes and indexes content, and supports multi-turn chat over a Milvus-backed knowledge base.
Hybrid retrieval with dense embeddings, BM25, and fusion.
Cross-encoder reranking for broad and ambiguous queries.
OpenAI-compatible SSE streaming API for Open WebUI integration.
Sentiment Analysis for Mental Health
PROJECT_2026
A comparative NLP system that evaluated classical baselines against a fine-tuned BERT model on a 40,000+ sample mental health dataset, then deployed the best production trade-off.
Benchmarked Naive Bayes and Linear SVC against fine-tuned BERT.
Reached 84.14% macro F1 with the selected BERT model.
Deployed the final model as a full-stack NLP web application.
LaTeX OCR
PROJECT_2025
A robust image-to-LaTeX system comparing six encoder-decoder architectures on the im2latex-100k dataset, with a Transformer decoder and beam search for stronger sequence generation.
Compared CNN, ResNet, and ViT encoder variants.
Achieved 91.46 BLEU-4 and 46.95% exact match.
Improved BLEU-4 by +14.6 over the additive-attention baseline.