Vivek Sharma

Who I Am

I'm a Machine Learning Engineer with hands-on experience architecting RAG systems and LLM inference pipelines that power real-world voice AI products.

My work sits at the intersection of applied ML research and production engineering — from building custom transformers from scratch to deploying scalable MLOps systems. I'm passionate about making AI genuinely useful.

Where I've Worked

Jr. Software Engineer — ML

Sep 2025 – May 2026

Anvex AI Technologies · Mumbai

Built RAG system for AnvexSpeak (AI Voice Agent) with multi-source extraction: client databases, Google Drive, web scraping; using BAAI/bge-large-en-v1.5 embeddings and advanced filtering.
Developed unified LLM inference layer using vLLM supporting Gemma 3, Qwen, Mistral, Llama and OpenAI models, with Qdrant (prod) and ChromaDB (dev) for document-grounded responses.

Machine Learning Intern

Jun 2025 – Aug 2025

Anvex AI Technologies · Mumbai

Built AVA — a RAG-powered policy chatbot using LangChain/LangGraph, Qdrant, and SambaNova Llama-4 across 26+ policies with FastAPI endpoints and confidence scoring.

Techathon Participant

Feb 2025

Mulund College of Commerce · Mumbai

Built ML models predicting equipment failures from sensor data and maintenance records, integrating with FastAPI for real-time predictive maintenance.

Things I've Built

MedVec-Scratch

Custom medical sentence embedding model built from scratch using a Siamese Transformer (4 layers, 256 dim) with a custom BPE tokenizer (30k vocab), trained on 233k medical triplets using Triplet Margin Loss for healthcare semantic search.

PythonPyTorchBPE TokenizerHuggingFaceTransformers

Federal Registry RAG Agent

Full-stack RAG system for US Federal Registry queries with Groq LLM, async pipelines, MySQL backend, and agentic tool-calling architecture for real-time document retrieval.

PythonGroq LLMMySQLStreamlitAsyncio

MLOps Pipeline with MLflow

End-to-end MLOps pipeline for medical cost prediction — comparing 4 regression models with automated best-model selection via R² score, MLflow experiment tracking, model registry, FastAPI serving, and Streamlit frontend.

PythonScikit-learnXGBoostMLflowFastAPI

ArtCycle

Deep learning project using CycleGAN to transform photos into paintings and vice versa — unsupervised image-to-image translation without paired data. Live on Streamlit.

PythonTensorFlowCycleGANStreamlit

NeuroScan

Streamlit web app classifying brain tumor types (Glioma, Meningioma, Pituitary, No Tumor) from MRI images using a pre-trained CNN. Live demo on Hugging Face Spaces.

PythonTensorFlow/KerasCNNStreamlitNumPy

Let's Connect

I'm open to ML engineering roles, research collaborations, and interesting projects. Drop a message!

VivekSharma