Welcome to my project portfolio! Here you’ll find a showcase of my Machine Learning and Data Science projects, each built with a focus on real-time data, interactive dashboards, and practical problem-solving.
Project 1
Agentic Market Intelligence Assistant (AgenticResearch)
PythonStreamlityfinanceRAGsentence-transformersGARCHLLM Agent
Description: An agentic market analyst that answers questions in plain English — “what’s driving NVDA today?”, “BTC-USD vs gold volatility?” — resolving every symbol named, pulling live quotes, history, news, and fundamentals from yfinance, running the numbers, and answering with sources.
- Live news + true RAG: headlines embedded with sentence-transformers and retrieved by cosine similarity across the news corpus and peer tickers — every story carries provider, date, and URL.
- Tool-using LLM analyst: agent can call get_quote (live price and day change) and rag_search (semantically retrieved headlines with source and date).
- Risk analytics powered by GARCH volatility forecasting on live market data.
Project 2
AmEx Credit-Card Default Prediction (AmexDefault)
PythonPyTorch1D-CNNLSTMGRUscikit-learnStreamlitGradio
Description: A credit-default prediction pipeline benchmarking 1D-CNN, LSTM, and GRU sequence models against a Random Forest baseline for 458,913 American Express customers, with a leakage-free pipeline and paired-bootstrap significance testing.
- On a held-out test set the GRU model reached PR-AUC 0.89 versus 0.87 for the baseline.
- K-Means clustering segmented customers into five risk tiers (default rates 2.6% to 93.2%).
- Live demo on Hugging Face Spaces (Gradio) + Streamlit app.
Links
🗡 GitHub
🔗 HF Space
🔗 Streamlit
🔗 https://huggingface.co/spaces/vivekkopthsd/amex-default-demo | https://amexdefault.streamlit.app/
🗡 https://github.com/bravo2024/AmexDefault
Project 3
Dynamic Flight Price Classifier – Real-Time ML Web App
PythonStreamlitscikit-learnPandasAmadeus APIMatplotlibSeaborn
Description: A real-time flight fare classifier using live Amadeus API data with pan-India coverage (30+ IATA city codes).
- Applied feature engineering (departure hour, weekday, trip duration, airline codes) and hyperparameter tuning.
- Boosted model performance from 60% baseline to 90–98% accuracy across routes (Kolkata↔Chennai: 96.8%, ROC-AUC: 0.995).
- Interactive Streamlit dashboard with custom route/date selection and 90-day fare simulations.
Links
🔗 Live App
🔗 https://flightpriceclassify.streamlit.app/
Project 4
TaxRAG | Indian Income Tax Q&A Assistant with Document Upload
PythonSentence-TransformersHugging Facescikit-learnStreamlitRAG
Description: A multi-turn Retrieval-Augmented Generation (RAG) chat assistant over the Indian Income-Tax Act, 2025, indexing 477 sections / 1,237 passages with section-level citations, hard abstention on out-of-scope queries, and user-PDF upload for private-document Q&A.
- Integrated a fine-tuned bge-small sentence-transformer bi-encoder reaching Recall@5 0.96, MRR@5 0.85 on a 100-query eval set.
- Backed by a multi-provider LLM chain (Groq, NVIDIA, OpenCode, Kilo) with automatic failover.
- Interactive Streamlit dashboard.
Project 5
Real-Time Market Price Forecasting Dashboard (Prophet + Streamlit)
PythonStreamlitProphet (Meta)yfinancePandasMatplotlibjoblib
Description: A real-time price forecasting dashboard for US/Indian markets supporting equities, indices, and cryptocurrencies.
- Integrated live yfinance API with Prophet for configurable forecast horizons (1–50 steps) across intraday to monthly intervals.
- Achieved low forecast errors: RELIANCE.NS (1D) – 0.91% MAPE; BTC-USD (daily) – 1.49% MAPE.
- Implemented model persistence to save/load trained Prophet models, reducing API calls and computation time.
Project 6
Movie Recommender System | Collaborative Filtering & Neural CF
PythonPyTorchscikit-learnPandasNumPyStreamlit
Description: A recommendation engine on the MovieLens-1M dataset (1M ratings, 6,040 users), implementing collaborative filtering, SVD matrix factorization, and a deep learning Neural Collaborative Filtering (NCF) model in PyTorch, evaluated under a leakage-free leave-one-out protocol.
- SVD performed best overall (NDCG@10 0.36, Recall@10 0.57, RMSE 1.15), ahead of NCF, item-item CF, and the popularity baseline.
- Interactive Streamlit dashboard for easy movie exploration.
Project 7
TxCat | Transaction Categorizer & Anomaly Detection
Pythonscikit-learnXGBoostTF-IDFIsolation ForestPandasStreamlit
Description: A text-classification system that categorizes 200,000 bank transaction descriptions into 10 spending categories using TF-IDF features with Logistic Regression and XGBoost, plus Isolation Forest for anomaly detection.
- Reached 97.5% accuracy, validated on a merchant-disjoint split (no merchant in both training and test) confirming generalization to new merchants rather than memorization.
- Added an Isolation Forest for anomaly detection.
- Interactive Streamlit dashboard.
Project 8
Human Activity Recognition (HAR) | Sequence Deep Learning
PythonLSTMGRU1D-CNNRandom ForestStreamlit
Description: A production-grade deep-learning pipeline for classifying human activities from smartphone inertial sensor signals, with an interactive Streamlit dashboard. Deep models (LSTM / GRU / 1D-CNN) are trained end-to-end on raw 9-channel × 128-timestep sensor windows, with a RandomForest classifier on 561 hand-engineered features as the classical baseline.
- Evaluation uses the official subject-independent train/test split (no reshuffling across subjects).
- Best model: 1D-CNN with 94.30% accuracy / 94.35% macro-F1, outperforming the engineered-feature baseline (RandomForest 92.87%).
- Interactive Streamlit dashboard.
Project 9
VisualProduct | Visual Retrieval & Multiclass Classification
PythonPyTorchscikit-learnXGBoostMobileNetV2Streamlit
Description: An end-to-end visual retrieval and multiclass classification system with Streamlit, demonstrating a retrieval pipeline that generalises across three benchmark datasets with different label structures: UT Zappos50K (50,025 images, 4 classes), Oxford-IIIT Pets (7,390, 37), and Caltech-101 (8,677, 101).
- Visual similarity search with three extractors: MobileNetV2 deep transfer features, PCA Eigenimages, and 3D HSV color histograms — sub-50 ms retrieval on CPU.
- Multiclass classification on MobileNetV2 embeddings, evaluated on a held-out 20% test partition.
- Interactive Streamlit dashboard.
Links
🗡 GitHub
🗡 https://github.com/bravo2024/VisualProduct
Project 10
CCMT M.Tech Cutoff Analysis & Prediction (2021–2025)
PythonStreamlitscikit-learnRandom ForestPandasPlotly
Description: An interactive Streamlit dashboard to analyze 5 years of CCMT M.Tech cutoffs and predict 2025 SR1, SR2, and NSR scores using Random Forest Regression. Users can upload CSVs (2021–2025), explore institute/program/category/round filters, and download prediction tables.
- Multi-tab UI: Upload & merge datasets, Explore raw tables, Cutoff trends, Round comparisons, 5-year institute history, and 2025 predictions.
- ML pipeline: RandomForestRegressor trained on 2021–2024; LabelEncoder for categorical features; auto-generation of SR1, SR2, NSR rows from 2025 R1–R3 inputs.
- Outputs: Downloadable CSV of 2025 predictions; clean filters for Institute, Program, Category, and Round.
Project 11
Fraud Detection AI Pipeline with Financial ROI Analysis
PythonXGBoostRandom ForestSMOTEStreamlitscikit-learnPlotly
Description: An end-to-end ML pipeline for credit card fraud detection on the European dataset (284,807 transactions, 0.17% fraud rate), with financial ROI analytics.
- Implemented SMOTE for class imbalance and compared 4 algorithms: Logistic Regression, Decision Tree, Random Forest, XGBoost.
- Achieved 94.3% recall with XGBoost, prioritizing fraud detection over accuracy (business-optimal tradeoff).
- Built an interactive financial ROI dashboard calculating dollar-value savings from fraud prevention vs. losses from false negatives.
Project 12
Customer Churn Prediction & Retention Budget Optimization (ChurnPredict)
Pythonscikit-learnXGBoostSHAPMixed-Integer ProgrammingSurvival AnalysisStreamlit
Description: A churn prediction and retention-budget optimization system over 1.76M e-commerce purchase records, aggregated to 5,013 customers × 24 features with temporal train/test cutoffs.
- Voting ensemble reaching ROC-AUC 0.898 and top-decile precision 0.487, with AUC 0.87–0.89 across 30–180-day churn windows and SHAP driver analysis.
- Retention budget allocated as a 0/1-knapsack MILP with HiGHS, plus survival analysis of time-to-return.
- A/B test power design for sizing future retention experiments.
Project 13
QuantRisk | Portfolio Risk Analytics Dashboard
PythonyfinanceVaR / CVaRPortfolio OptimizationFactor ModelsMonte CarloStreamlit
Description: Live NSE, US and global portfolio risk analytics powered by yfinance, with benchmark alpha, tracking error and information ratio.
- 95% / 99% VaR and Expected Shortfall computed by historical, parametric and Monte Carlo methods, plus drawdown and per-asset risk contribution.
- Fama–French 5-factor regression and an efficient frontier of max-Sharpe, minimum-volatility and risk-parity portfolios.
- Fully interactive Streamlit dashboard with live data refresh.
Project 14
MRSS | Mortgage Risk Segmentation & Survival Analysis
PythonPCAK-MeansCox PHSurvival Analysisscikit-learnStreamlit
Description: Risk segmentation and survival analysis over 62,359 FHFA mortgages (2013–2024) on 17 origination features.
- Segmented with PCA + K-Means and cross-checked by Ward clustering, recovering a prime tier (FICO 773, LTV 56%) and a higher-risk tier (FICO 734, LTV 84%).
- Kaplan–Meier, log-rank and Cox PH on 90-day delinquency: C-index 0.858, rising to 0.902 on 2021–2024 out-of-time loans.
- Interactive Streamlit dashboard for tier exploration and survival curves.
Project 15
BookRec | Goodbooks-10k Recommender & Semantic Search
PythonSentence-TransformersPyTorchRecommender SystemsSemantic SearchStreamlit
Description: A recommender and search engine over 10,000 books, 5.98M ratings and 53,424 readers.
- Five methods compared: semantic content search, sparse SVD, item-item CF, Neural CF and a hybrid, with explicit cold-start handling.
- Best model (item-item cosine CF): NDCG@10 0.37, Recall@10 0.56, RMSE 0.91.
- Interactive Streamlit app for search and recommendations.
Project 16
IEEE-CIS Fraud Detection & Anomaly Risk Monitor (IEEECIS)
PythonXGBoostGraph FeaturesCalibrationIsolation ForestStreamlit
Description: Fraud detection over 590,540 transactions (3.5% fraud) on an out-of-time split, with 137 engineered entity-history and PCA features.
- XGBoost: PR-AUC 0.477, ROC-AUC 0.881, Precision@1% 0.874; isotonic calibration cuts the Brier score by 33%.
- Graph fraud-ring features lift PR-AUC to 0.566, while an Isolation Forest anomaly layer stays stable under refits (Jaccard 0.64).
- Interactive Streamlit risk monitor.
Project 17
FraudGuard | Credit Card Fraud Detection Under Extreme Imbalance
PythonLightGBMClass ImbalanceCost-Sensitive LearningCalibration
Description: Fraud detection on the ULB dataset: 284,807 transactions with only 0.17% fraud.
- LightGBM: ROC-AUC 0.979 and PR-AUC 0.845 on 71,201 held-out transactions.
- Cost-matrix threshold tuning and calibration, with an optional chronological split.
Links
🗡 GitHub
🗡 https://github.com/bravo2024/FraudGuard
Project 18
PcamVLM | Histopathology Vision-Language Model
QLoRA / PEFTVision-Language ModelsMedical ImagingTRLPyTorch
Description: A QLoRA fine-tune of MedGemma 1.5 4B (4-bit NF4, LoRA r=16) trained with the TRL SFTTrainer for histopathology classification.
- PatchCamelyon: accuracy 0.908, sensitivity 0.872, specificity 0.943; 9-class NCT-CRC: accuracy 0.876.
- Adapter published on Hugging Face with a Streamlit demo.