A production-grade, multi-agent financial intelligence system that answers natural-language stock queries using deterministic fundamental and technical analysis. The platform dynamically ingests missing symbols, caches structured micro-indexes, and allows users to choose between existing or freshly fetched market data, ensuring accuracy, transparency, and auditability through a unified API and interactive UI.

Tech Stack:
Python 3.11, pandas, NumPy · Alpha Vantage API · Rule-based Technical Indicators (RSI, SMA, EMA) · CSV Micro-Indexes · Gradio UI · Modular Agent Architecture · File-based Caching & Locking

This project implements an end-to-end computer vision pipeline for automated skin lesion classification using the ISIC 2018 dermoscopic dataset. Two approaches were evaluated: a custom CNN trained from scratch and a transfer learning model based on EfficientNet-B3. The system supports both 7-class diagnosis and binary malignant vs benign classification.

Model performance was evaluated using standard metrics including accuracy, precision, recall, F1-score, ROC-AUC, and confusion matrices. Grad-CAM was integrated to provide visual explainability by highlighting clinically relevant regions.

Tech Stack: Python, PyTorch, EfficientNet, OpenCV, scikit-learn, Grad-CAM, Gradio

Features

• End-to-end pipeline for extracting entities from .tif and .pdf legal documents
• Supports multiple NER models: spaCy, BERT, Legal-BERT, and Regex
• Fully offline processing — no cloud or third-party APIs involved
• Outputs in .json or .csv with confidence scores
• UI built with Gradio; Regex extractor also available via Streamlit

Tech Stack 

• Frontend UI: Gradio, Streamlit
• OCR Engines:   Tesseract, PaddleOCR 
• NER Engines:    spaCy, Bert, Regex 
• Data Handling:  pandas, re            
• Deployment: Local execution or cloud

• LSTM-based time series forecasting
• Interactive Streamlit UI with slider-controlled hyper-parameters
• Reproducible training with adjustable epochs, window size, batch size, dropout, etc.
• Live visualization of training & validation loss
• Real-time prediction vs actual price plotting
• Automatically loads pretrained model for demo on app launch

TECH_STACK: Python 3.11, TensorFlow / Keras , Pandas, NumPy,Matplotlib, Seaborn ,LSTM Neural Network, .h5 (Keras model serialization format), Wandb (Weights & Biases) for experiment tracking. Deployment: Streamlit , Hugging Face Spaces, Version Control, Git & GitHub, miniconda

An AI-powered text analytics pipeline designed to help financial analysts & compliance teams process high volumes of textual analyst commentary, identify if the commentary is Positive / Neutral / Negative,determine whether it contains risky financial signals, and extract key entities like company names or financial instruments — all through a single, unified API.

Tech Stack:

• Language & ML: Python 3.11, scikit-learn, XGBoost , Data Processing: pandas, numpy, CountVectorizer
• NER: Azure Cognitive Services (Language API) . Model Serving: FastAPI, Uvicorn
• Deployment: Docker, Azure Container Registry, Azure App Service
• Versioning & Testing: joblib, PyTest .Visualization (for EDA phase): matplotlib, seaborn (Jupyter)
• GitHub:https://github.com/ecubeproject/Equity_Risk_And_Sentiment_Intelligence_Engine

Features

VAR-based multivariate time series forecasting. Choose forecasting horizon from 4 to 12 weeks. Select any combination of assets: Bitcoin, Gold, and S&P 500

Forecast plot and Forecast table available for export. Built with Gradio UI for interactivity

Tech Stack

• This project is an end-to-end machine learning pipeline to predict the Click-Through Rate (CTR) of marketing campaigns using a hybrid recommendation approach. 
• It combines XGBoost regression with a Streamlit UI and is deployed on Google Cloud Run.
• Business Objective: Given historical campaign data, predict the CTR to help marketing teams optimize campaign performance and ROI.
• Tech Stack:

• Python 3.9, scikit-learn, XGBoost, Streamlit, MLflow
• EDA & Visualization: matplotlib, seaborn
• Google Cloud Run, Cloud Build, Artifact Registry

Campaign Success Prediction – End-to-End ML Deployment on GCP Vertex AI

• Business Problem: A marketing agency runs thousands of digital ad campaigns. The business wants to predict whether a campaign will succeed (e.g., meet conversion targets) based on parameters like budget, platform, ad type, CTR, and audience segment.
• Goal: Help campaign managers estimate success upfront — to optimize ad spending and maximize ROI.
• Data Science Framing

• We re-framed this into a binary classification problem
  • Target: Whether a campaign is successful (1) or not (0)
  • Features: Budget, Impressions, CTR, Past Success Rate, Ad Type, Target Age Group, etc

• This is the Gradio based prototype of a smart parking solution Kivy mobile app that we presented in a local hackathon and won second prize. 
• A high quality research backed dataset-considered benchmark in computer vision was used. (Source)
• The prediction model was made using Vision Transformer and it outclassed all other algorithm showing that transformer architecture, which was originally designed for NLP works equally well for image classification.
• Final app had enhanced version of this including object detection, image detection, ability to handle streaming data, QR code based authentication for end users. 

• Stroke is the second largest cause of death, according to WHO. The image shows devastating impact of stroke
• However, stroke can be easily prevented if we can successfully identify at risk patients by easily monitoring simple clinical and life-style features. 
• But for prevention, first we need to accurately predict the stroke.
• So, we designed this app to predict stroke with 95% precision, recall and F-1 score and astounding 99% AUC-ROC on actual patients' data. The analysis is based on an excellent research paper in the reputed journal nature.
• A functional streamlit app built by us is deployed here.

• This is loan eligibility chatbot based on real data of a US financial institution.
• User enters various options and a profile of user is built. 
• Instantly ‘Loan Approved’ or ‘Loan Not Approved’ decision is displayed with a message and image.
• User gets reasons which factors of his profile are helping approval and which are pulling against approval.
• The highlights of the app are its lightweight footprint which leads to real time decision and interpretablity of results.
• A very accurate and efficient random forest model. 

• This is a heart attack prediction app based on real data and easy features so that even a layman can do self-health check.
• Features are Gender, Age, CigarettesPerDay, BPMeds , PrevalentHypertension, Diabetes, TotalCholesterol, SystolicBP, DiastolicBP, BMI, HeartRate, FastingBloodGlucose.
• User enters value and he can see probability of heart-attack. A custom cut off of 42% was used to get the best F1-Score so as to minimize both false positives and false negatives to the most optimal level.
• A powerful XGBoost model with Grid search CV and 5 fold cross validation , having F1- score 96.7% was obtained for very precise diagnosis.
• Post-diagnosis, next stage is advising patient on what are the most important risk factors for him/her. Here EACH PATIENT IS DIFFERENT. So, it is possible to see which factors are pulling the prediction to 1 heart-attack (red) and which are pulling to 0 no heart-attack. (blue). A new diagram is generated for each patient based on values entered by him.

• This is a highly scalable customer churn application using real data (10000 rows)  of a multinational European bank.
• The business problem is:

1.  To predict which customers are likely to churn based on customer's features and history.
2. To find mean and median churn probability for  each feature-group for  the high churn list.
3. To study which factors are causing churn and which are preventing churn for an individual customer.
4. To develop a Flask app with JavaScript, HTML, CSS and trained XGB model and deploy on AWS.

• Data science problem is  to do binary classification, to create a reproducible pipeline , try different algorithms and find the best algorithm. Extensive data preprocessing and feature engineering are involved.
• The application was deployed on AWS using state of the art MLOps technique.
• Key features of the application are reproducibility, interpretability and explainability due to thoughtfully designed user interface which makes things clear in just few clicks by attractive and informative plots.

• This is a standalone python app with minimal footprint but great features.
• User uploads CSV or XLSX data file as input. Size of the data set is limited only by user's hardware capacity. 
• App displays necessary EDA information and statistical summary.
• User selects his response variable for binary or multi-class classification. 
• App allows user to select one of 13 classification algorithms. 
• User adjusts hyper-parameters for chosen classifier by convenient controls. 
• For each classifier and each set of hyper parameters the model training and Evaluation are performed independently and evaluation metrics and classification reports are seen in output-2. During training data pre-processing steps are also performed in the pipeline meaning that input data need to be 100% clean.
• Once the best classifier and best hyper parameters are chosen based on characteristics of data and the use-case, the user can plot seven different types of plots.
• Below each plot helpful short information is printed explaining how to use that plot for interpretation and inference.
• The application be easily ported to be deployed on any OS or on web.

• This is the POC version of a real life project. It is a multi-page python flask application which can be deployed on any cloud. 
• Real and recent price data of six Indian metros (Mumbai, Delhi, Kolkata, Chennai, Hyderabad and Bengaluru) was used.
• As data distribution for each city was different, six different pickle models - one for each city- were prepared.
• As location was affecting price in same city, so, location was kept as feature. However, presence of hundreds of location made it challenging to design the app and subsequent testing.
• Many regression algorithms were tried  before deciding on XGBOOST with  5 fold cross validation and grid search CV to decide on the optimal hyper parameters.
• Impressive evaluation metrics were obtained for both train and test set. Even in actual production environment performance was very good. 
• Later, on client's request, new features like RERA approval status, rental yield and a recommendation systems.were added.

• This is POC  version of E2E ML project done by us for e-commerce portal. 
• Business requirement was to develop and deploy an accurate email spam-ham detection model with provision for automatic model drift detection and model monitoring.
• Tools used were streamlit, MLFlow, Apache Airflow,Docker, Evidently AI.
• We developed this model to maximize  precision and F1 and also focus more on  increasing AUCPR due to imbalanced data.
• We also handled deployment,  automated scheduled model drift analysis and model retraining using Airflow and DAG..

• This is a POC voice transcribe app that uses open AI whisper API at back-end.
• User selects an audio file on his PC and the app generates text transcription for audio
• The audio language is automatically detected and transcription is in same language.
• This can have application in medical transcription.

• This is a minimal tkinter based language translation app using openAI API at the backend.
• It allows user to browse for a MS-WORDS (.docx) file on his PC. The file must contain English text.
• User then selects target language and at the click of a button the translated text in that language is visible in the text box on the UI of App

• This is a simple image generator application using open AI API and DALLE-3.
• A minimal user-interface using tkinter library is provided with a text box where user types his prompts and another window for output image.
• We have developed its commercial version for clients in textile sector using Leonardo AI API which is favorite for textile designers.

• This is a streamlit app that allows user to enter no of slide that he desires to create in presentation.
• It next dynamically generates as many text fields as the number of slides for entering prompts.
• User enters prompts for each slide in this prompt fields
• On pressing create slides button presentation is generated with text and corresponding images created by Dalle.

• This is an essay generator app using PyQT6 to build the UI and open AI API at back-end to generate essay on any topic.
• At the click of a button essay of desired length on desired subject will be generated and at the click.
• We are have deployed a refined version of this app with attractive UI in a few educational institutions.

• This is a much more comprehensive Django based Quiz generator application.Django framework is used for apps like Instagram or pinterest which is testimony to its robustness.
• App can generate quiz on any topic provided as text input in its input window.
• The generated questions, along with answers, are displayed in output window.
• The generated quiz is also saved as text file and a database.

• This is a  programming language agnostic code debugger which takes as input a wrong code  and gives corrected code in output.
• It also gives a detailed explanation of the error. Making it a great programming language teaching tool. 
• We actively use it in our firm to train interns in coding assignments.
• The app can also be easily deployed on Azure cloud and integrated with a payment system for commercial use.

• This is a python flask project that works as chatGPT clone
• This app works as stand alone application and it can be used as chatGPT clone
• The user needs to have chatGPT API key (i.e., paid subscription).
• Best practices are followed by providing a separate config file to store API key rather than hard-coding it inside application. 
• The screenshot of running app shows its successful working. 

• This is POC streamlit app that aids physicians to arrive at a diagnosis about probability of a patient having diabetes using easily measurable features.
• Although a real dataset was used in predictive model building, the data is old and of single ethnicity. So, the app is used to train interns in our firm and by our marketing team to show POC when they visit clients in health-care sector. 
• Future scope: To integrate the app  with wearable devices and send the data to physician's dashboard. 

• This is a streamlit app that performs multi-class classification using Naive-Bayes algorithm. Originally made for interns of my firm to teach them machine learning model building, deployment and front end programming, the app won second prize in a reputed hackathon.
• This is a crop recommender system. It takes seven features as follows as input and suggests the best crop to grow for that soil condition and climate:  N – Ratio of Nitrogen content in soil,  P – Ratio of Phosphorous content in soil,  K – Ratio of Potassium content in soil, Temperature – temperature in degree Celsius,  Humidity – relative humidity in %,  ph – ph value of the soil, Rainfall – rainfall in mm, 
• This app can be integrated with IOT sensors and , in green houses, even the climate can be controlled to get bumper harvest of desired crop.