Jayant Kothari

I’m an AI/ML engineer focused on building production-minded NLP and generative AI systems. I enjoy turning research ideas into end-to-end pipelines that are measurable, fast, and maintainable.

In my ML internship, I built an end-to-end analytics platform with XGBoost to predict academic performance, attendance, and institutional health across 15+ partner schools—achieving 87% accuracy and 0.91 ROC-AUC over 10,000+ records. I automated evaluation dashboards, cutting manual effort by 60% and reducing report turnaround from 3 days to 4 hours. I also deployed inference via FastAPI microservices with p95 latency under 200ms.

I’ve been publishing as well—my CHANDICON 2026 work systematically compares Transformer adaptation families (FFT, PEFT, LoRA, RLHF), benchmarks them on GLUE with RoBERTa-base, and identifies key research gaps like alignment robustness and evaluation standardisation. I’m especially interested in practical fine-tuning strategies that preserve quality while staying resource-efficient.

Across my projects, I focus on retrieval-augmented reasoning and real-time delivery: DocuMind combines asynchronous multi-stage RAG with hierarchical chunking and hybrid retrieval, handling 1,000+ page PDFs while cutting redundant LLM calls by 40% and query latency by 55%. I’ve also built a real-time fraud-decisioning microservice using supervised/unsupervised fusion with SHAP explanations, and a multilingual YouTube smart chatbot that supports long videos with timestamp-synced, grounded responses.