Mayank Joshi

A lightweight LLM inference engine built from scratch in PyTorch. Supports Llama, Qwen, and Mistral models with 4-bit NF4 quantization (bitsandbytes), FlashAttention-2, and stateless sampling.

A civic-tech AI platform that transforms multilingual citizen reports from public feeds and APIs into geospatially-aware infrastructure proposals using sentence embeddings, HDBSCAN, and LLM agents.

A high-precision word-matching engine for low-resource translation. Uses multilingual sentence and token embeddings to perform mutual-consent token alignment and extract bilingual dictionaries.

AI-powered data platform for oceanographic NetCDF datasets. Leverages a RAG architecture to support natural language-to-SQL query generation across 580+ NetCDF profiles.

The first installable Kangri-Hindi machine translation package for Argos Translate. Trained a 6-layer Transformer from scratch on 26,779 parallel pairs and deployed using CTranslate2.

An AI-powered visual inspection system with 90% faster quality analysis. Combines a convolutional feature extractor with a ConvLSTM temporal module and Streamlit feedback loop.