View on GitHub

usama_Portfolio

Data Science Portfolio

      *click on the project title to view the code*

We have created a multi-agent chat model that allows you to perform various tasks, such as sending emails to people or browsing the web
We have utilized Langchain and Langgraph, which extends the LangChain Expression Language with the ability to coordinate multiple chains (or actors) across multiple steps of computation in a cyclic manner
A supervisor will understand the task and utilize one or more agents, depending on whether they can answer or solve the desired task
Once an agent completes its task, it will provide the result to the supervisor, deciding whether the task is completed or if another tool is required. Regardless of the situation, the chain will come back to the supervisor

Understanding the mechanics behind LoRA, QLoRA, and prompt tuning.
Step-by-step guides on applying these techniques for fine-tuning.
Evaluating the performance and efficiency of PEFT techniques against traditional fine-tuning approaches. Repository Structure
Demonstrates prompt tuning on Llama-2-7B for targeted NLP tasks.
Showcases the application of QLoRA for fine-tuning on specific datasets.
A PowerPoint presentation detailing our experiments’ PEFT concepts, methodologies, and key findings.
Demonstrated the prototype of LLM Fine-tuning (PEFT) as-a-service where you can upload the data, train multiple models and decide which model to use based on Rouge and BERT scores (Note: The code for this prototype is not included in the repository as the company owns it and cannot be shared publicly.)

Cleaned the text data by removing stopwords and lemmatization
Performed EDA using Word Cloud and seaborn to analyse the word’s effect on reviews
Performed Layer Embedding on the text data and created a model using LSTM after comparing the different models with accuracy of 85%
Integrate the model on Flask and deploy it on Heroku
https://women-clothing-review.herokuapp.com/

Deal with missing value in ZONE using Random Forest
Cluster the data with the help of state, zone and group
Modeling and predicting the application using a Decision tree after comparison on root mean square error with a score of 469.3362715213254
Also used forecasting to predict Application count for the next 3 months

Scrap the Glassdoor data using Python
Cleaned the data using R
Performed EDA and extracted some valuable insights like the job availability, skills required, companies where jobs are available in COVID-19, etc

Deal with the null value with the help of the imputer method in Python
Analyze the data to derive different insights and help to choose the algorithm
Compare different algorithms and selected Extra Trees Regressor based on a negative mean absolute error with a score of -753.9840726814722

Cleaned and preprocessed the data using NumPy, Pandas, seaborn
Reduced the dimensionality using PCA
Modeling and Prediction of trainee performance using a Decision Tree with an accuracy of 80% by F1 Score

Cleaned and analysed the data sets using python
Using ARIMA and Exponential smoothing fitted and forecasted the non-seasonal and seasonal variables respectively

Cleaned the data using Python library pandas
Analysis of different symptoms of the virus, the origin, spread of the virus and Dashboarding using Tableau

Cleaned the data using Python libraries like Numpy and pandas
Performed EDA like top apps, types of preferred paid apps, etc and statistical analysis

This Project contain Analysis and Dashboarding in Excel
PPT of Analysis of Indian Premier League from 2008-2019 for Top 10 Batsmen and bowlers, teams which are difficult to win against, Stadiums preferred for batting and bowling, etc