AI Lead at AllNerds driving full cycle delivery of scalable agentic products Expert in RAG,LLMs and cloud native systems

Ankush Mulkar  • Peerlist

Developing a question-answer bot that leverages Amazon SageMaker, Amazon OpenSearch Service, Streamlit, LangChain, and FastAPI to provide users with accurate answers to their questions using a knowledge corpus.

Developing question answering bot with Amazon

The objective of this project is to create a chatbot using the Llama 2 language model powered by Replicate. The chatbot will be hosted on a Streamlit app and will interact with users in real time, providing responses based on the input it receives. The chatbot aims to assist users by answering their queries effectively and efficiently

Chatbot using Llama 2

The objective of this project is to create a chatbot that can answer technical questions about the Streamlit Python library. The chatbot leverages LlamaIndex and OpenAI's GPT-3.5-turbo to provide accurate and technical responses based on the indexed Streamlit documentation.

Llamaindex-chat-with-your-documents

This Streamlit app allows you to generate blog content using the Falcon-7B Instruct language model. Simply provide a topic, and the app will generate a blog with a detailed introduction, explanation, and summary about the topic.

Blog-generator-using-falcon-7b

"Ask Buddy" is an interactive chatbot application built using Streamlit and integrated with the HugChat library. The chatbot aims to provide real-time responses to user queries and inquiries, making it a valuable tool for answering questions and engaging in conversations.

Ask Buddy - Streamlit Chatbot Application

This project is to predict fraudulent credit card transactions with the help of machine learning models. In this project, you will analyze customer-level data that has been collected and analyzed during a research collaboration of Worldline and the Machine Learning Group. We balanced the data with various approaches such as Undersampling, Oversampling, SMOTE, and Adasy. With every data balancing technique, we built several models such as Logistic, XGBoost, Decision Tree, and Random Forest. the SMOTE and Adasyn models performed well. Among those models, the simple model Logistic regression has an ROC score of 0.99 in the train set and 0.97 on the test set. we can conclude that the Logistic regression model with SMOTE is the best model for its simplicity and less resource requirement.

Credit card fraud detection

Emplyeee_analytics-Employee will churn or not? Deployed on Streamlit app https://ankushmulkar-emplyeee-analytics-stramlit-v991uw.streamlit.app/

Employeee analytics-Employee will churn or no

https://github.com/AnkushMulkar/Sentiment-analyser-app-Hugging-face-transformers-Streamlit This application classifies a user's input as POSITIVE or NEGATIVE based on the default Hugging Face Transformer model. The web app is built using Streamlit and can be deployed using Streamlit Sharing. click the below link to visit the deployed model- https://ankushmulkar-sentiment-analyser-app-h-sentiment-analyser-v3gmyy.streamlit.app/

Sentiment analyzer app Hugging face transform

To build a system that categorizes leads based on the likelihood of their purchasing CodePro’s course. This system will help remove the inefficiency caused by junk leads in the sales process. CodePro is an EdTech startup that had a phenomenal seed A funding round. It used the money to increase its brand awareness. As the marketing spend increased, it got several leads from different sources. Although it had spent significant money on acquiring customers, it had to be profitable in the long run to sustain the business. The main objectives of lead scoring are as follows: Remove Junk by categorising leads on the basis of propensity to purchase Gain insights to streamline lead conversion and address improper targeting We have chosen L2AC (Leads to Application Completion) as our business metric, as choosing L2P (Leads to Payment) will aggressively drop the leads. A lead is generated when any person visits CodePro’s website and enters their contact details on the platform. A junk lead is generated when a person who shares their contact details has no interest in the product/service. Having junk leads in the pipeline creates significant inefficiency in the sales process. we are creating three different pipelines for our use case. Data Pipeline Training Pipeline Inference Pipeline Our system metrics for the ML model are : AUC score >75% Precision > 65% Recall > 75% Tools used- Airflow,Mlflow,Pycaret,Pandas profiling

Mlops Lead scoring for edutech company

In this case, study, imagine I was working as an NLP engineer for a financial company that wants to automate its customer support tickets system. As a financial company, the firm has many products and services such as credit cards, banking, and mortgages/loans. For a financial company, customer complaints carry a lot of importance, as they are often an indicator of the shortcomings in their products and services. If these complaints are resolved efficiently in time, they can bring down customer dissatisfaction to a minimum and retain them with stronger loyalty. This also gives them an idea of how to continuously improve their services to attract more customers. These customer complaints are unstructured text data; so, traditionally, companies need to allocate the task of evaluating and assigning each ticket to the relevant department to multiple support employees. This becomes tedious as the company grows and has a large customer base.

Automatic Ticket Classification

dentifying entities in Healthcare data You need to process and modify the data into sentence format. This step has to be done for the 'train_sent' and ‘train_label’ datasets and for test datasets as well. After that, you need to define the features to build the CRF model. Then, you need to apply these features in each sentence of the train and the test dataset to get the feature values. Once the features are computed, you need to define the target variable and then build the CRF model. Then, you need to perform the evaluation using a test data set. After that, you need to create a dictionary in which diseases are keys and treatments are values.