At Planto AI, I optimized Retrieval-Augmented Generation pipelines with re-ranking and semantic chunking, increasing LLM accuracy by 18% and reducing latency by 25%.
I also led a team of five engineers in shipping an AI-powered VS Code extension using multi-agent workflows, and engineered Hugging Face Transformers and PEFT (LoRA) fine-tuning workflows that reduced cloud infrastructure costs by 40%.
On freelance projects, I built an offline Flutter mobile AI system for rice quality classification on low-end Android devices. I optimized a ConvNeXt Vision Transformer with ONNX graph surgery and FP16 quantization, reducing its memory footprint by 50%.
I also developed a satellite time-series system to classify crop types from Sentinel-2 imagery, combining PatchTST transformer embeddings and vegetation indices with a soft-voting ensemble. My projects include a multi-agent financial news system and a medical LLM fine-tuned for clinical question answering.
