At MAYA Data Privacy, I replaced LLM inference with a CPU-optimized model, reducing latency from two minutes to one second per row. I also improved PII detection accuracy to 95% with span-based NER.
I redesigned APIs as an Associate Developer Intern at MAYA Data Privacy, moving them to a job-based asynchronous model and improving uptime to 99.9%+. I also improved detection accuracy by replacing fixed-size chunking with context-aware semantic chunking.
At Western University, I built a hand-gesture recognition system aimed at minimizing driver distraction, with real-time inference under 100 ms. The seven-way, five-shot system achieved 99.87% accuracy using PyTorch.
At Ennoventure, I curated AI-generated and authentic video and audio datasets to enhance DeepFake detection, and developed benchmarking pipelines for more than 20 pre-trained models. My projects include KrishiNext, a multilingual agricultural knowledge assistant using RAG, FAISS, and vLLM.
