Bitcoin Price Predictor

Deep Learning model is used to predict the next day close price of Bitcoin. Input to the model was OHLCV data fetched using CoinCapmarket or Bitfinex API

Python

Call Centre Analytics using NVIDIA Stack

E2E Conversational AI solution for a global call centre. Developed pipelines to transcribe, analyse & categories customer interactions, enabling sentiment analysis, agent performance insight and post-call analytics. RAG was used for non-parametric knowledge.

RivaTritonGKENIMsBertRAGAgents

ConstiQA

LLM powered chatbot that was developed to QnA - “The Constitution of India”.RAG based approach was followed using LlamaIndex.

LLamaindexMistralOllama

Contract Complaince & Legal Analysis

GenAI powered application for effective Legal Contract Compliance and Analysis. RAG pipelines were designed for storing chunked data into VectorDB (Knowledge Base) and for generation. Self-RAG technique was utilized. Evaluation method powered by Human-in the-loop. SLM model was used for local deployment using Ollama.

LangchainNeo4jRAGHuggingfaceLangGraph

Fine-Tuning a Reasoning Model using NVIDIA Stack

Researched and Fine-tuned SoTA Reasoning model for Domain-adapted Legal Language Understandng. NVIDIA's Nemo microservices are utilized under DataFlywheel settings.

NIMsNemoGitopsCUAD Dataset

Char-RNN based Gita-Generator

A fun project to generate verses of BHAGVAD GITA using Neural Networks

TensorFlowPythonGoogle Colab

HumSafar

This LLM based Chatbot designed to provide support to elderly people who needs care assistance.Open-sourceframework was used to develop this Telegram Bot. Mistral 7B model was utilized to generate responses.

OpenAILangChainRasa

KGXplorer

Developed as an interface to data stored in DW/Knowledge Graphs for insurance fraud detection.Langchain was used as orchestration framework along with neo4j as graph datastore. Vertex AI and GCP provides pipeline to feed the KG from data stored in BigQuery.

Neo4jLangchainVertexAIGCPBigQuery

News Verasity Validation using GNN

Classification model to determine News article veracity was implemented using Graph Neural Networrks.Both the structural and language modeling aspects were taken into consideration. PyG is used to analyse the propogation patterns and graph attention that would compliment and supplementthe language model.BERT language model was used to generate textual (tweets) representation alogng with some pre-determined profile features. Triton Inference server was utilized to expedite inferencing on cloud-based deployment.

PyTorch GeometricPytorchTriton

Time-Series Prediction using Transformers

This research project was implemented to assess the applicability of 'blue-eyed' model of langauge world to TS space. Synthetic data generation capabilities of transformer were explored.

LLMsGenAISynthetic Data

Talk2DB

The idea is to interact with the database using Natural Language Processing commands. OpenAI’s model is used to generate embeddings and SingleStore act as vectorstore. Langchain’s capabilities were utilized to generate SQL code & NLP commands, both closed source and open-source options were explored.

OpenAILangchainSingleStore

Vehicle Licence Plate detection using Deep Learning

TensorFlowPythonOpenCV