Projects
// GenAI & NLP
UK Visa RAG Chatbot
RAG chatbot that ingests UK Immigration Rules via GOV.UK API, embeds them with sentence-transformers (all-MiniLM-L6-v2), stores in FAISS/ChromaDB, and generates answers with source attribution through a Streamlit chat interface.
Local Peer Review
Automated academic peer-review generator that verifies citations, runs adversarial probes, and produces structured referee reports. Supports Ollama, OpenAI, Groq, and LM Studio backends with a Gradio web interface.
Resume Customizer
Automatically tailors LaTeX resumes to match job descriptions using the Google Gemini API, with structured output parsing and PDF generation.
LaTeX for LLM
Converts LaTeX documents into a typed graph for targeted LLM context retrieval, where graph-based queries cut prompt size by ~54% on average (up to ~80% for focused queries). Features cross-reference resolution, inline math preservation, interactive force-directed graph visualization, and a Claude MCP server.
arXiv RSS Filter
Streamlit app that filters and curates arXiv preprint RSS feeds, helping researchers stay on top of relevant new papers in their field.
// Machine Learning
Seoul Bike Demand MLOps Pipeline
End-to-end, reproducible ML system for hourly bike-demand forecasting, built to run entirely on a laptop. A DVC pipeline chains ingestion, Pandera validation, feature engineering, LightGBM training, and evaluation; MLflow tracks every run and a gated registry only promotes a new champion when holdout RMSE improves. The champion is exported and served through a Dockerized FastAPI app, with Evidently drift monitoring wired to a retrain signal and CI on GitHub Actions. Holdout R² ≈ 0.80 on a time-ordered split.
Galaxy Morphology Deep Learning
PyTorch image classification of galaxy morphology on Galaxy10 SDSS: 21,785 real SDSS cutouts across 10 heavily imbalanced classes. Benchmarks a from-scratch CNN against ResNet-18 transfer learning, both frozen-feature extraction and full fine-tuning. Class-weighted loss and macro-F1 model selection handle the imbalance honestly; the fine-tuned ResNet-18 reaches ≈ 0.80 test accuracy and 0.64 macro-F1. MLflow tracks every run and a Dockerized FastAPI service returns per-class predictions, with CI on GitHub Actions.
UK Accident Severity Classification
Dual-strategy LightGBM classifiers on 104K DfT STATS19 collisions merged with 190K vehicle records (42 features). Severe-optimized model (0.527 macro recall, 66.8% accuracy) with threshold tuning for adjustable recall-accuracy tradeoff. Results benchmarked against published studies on the same dataset.
M25 Congestion Forecasting with LSTM
PyTorch LSTM that forecasts motorway congestion from real Transport for London sensor data. Each checkpoint reports traffic volume and average speed every 15 minutes; the model reads a 4-hour window of history and predicts whether the next interval will be congested (average speed below 30 mph). Congestion is the minority class at this site (≈ 10% of all intervals), and because the 70/15/15 split is chronological its rate climbs from 5.5% in training to 25.6% in the held-out test period — so the model is benchmarked against majority-class, rush-hour and speed-threshold baselines rather than judged on accuracy alone: 0.977 recall and 0.765 F1 on 15,731 held-out windows, with PR-AUC 0.893 against a no-skill baseline of 0.256, where always predicting “not congested” scores 0.744 accuracy but 0.000 F1.
Deep Reinforcement Learning
Implementations from Deep RL in Action covering policy gradients, Q-learning, and multi-agent environments using PyTorch.
// Quantitative Finance
Risk Management Toolkit
End-to-end risk analytics pipeline: raw price data is denoised with Kalman filters and wavelet transforms, fed into technical indicators (MACD, Bollinger, PSAR, stochastic oscillators), then evaluated via historical & parametric VaR, CAPM beta, and Sharpe/Treynor/Jensen performance metrics.
Time Series Forecasting
Comparative equity modeling: ARIMA for trend forecasting, GARCH for volatility clustering, CCC-GARCH for multi-asset portfolio-level forecasting, and Prophet for seasonality. Includes rolling & fixed window evaluation, bootstrap simulation, and model diagnostics on Yahoo Finance data.
Backtesting Framework
Strategy backtesting engine built on VectorBT's vectorized execution, progressing from moving-average crossover strategies to volatility-regime-aware ML signals incorporating GARCH-estimated conditional variance.
ML Trading Signals
Momentum-based trading signal generation using MA ribbons, distance metrics, correlation-based signals, and physics-inspired momentum measurements.
AI for Trading
Udacity nanodegree projects: momentum strategy, breakout strategy, and smart beta portfolio optimization with performance evaluation.
Mutual Fund Analyzer
Live Streamlit Cloud app that ingests Morningstar fund data to build multi-fund portfolios and analyze them across sectoral distribution, scheme comparison, valuations, and top company exposure.
NSE Stocks Dashboard
Streamlit dashboard fetching live data from the NSE API for real-time stock analysis, price monitoring, and technical charting of equities listed on the National Stock Exchange of India.
// Astrophysics Research
AGN & Local Environment in HR5
Investigated effects of local environment on Active Galactic Nuclei activity using the Horizon Run 5 cosmological simulation with Python-based feature extraction pipelines on terabyte-scale data.
Dust Radiative Transfer (HR5)
Modified the Powderday radiative transfer code for Horizon Run 5 AMR data. Built pipelines for galaxy selection, sub-cube extraction from HDF5 snapshots, and SED computation with MPI parallelization.
Cosmic Web Analysis at Gigaparsec Scales
Developed a multi-block approach for the DisPerSE cosmic web finder enabling analysis at unprecedented gigaparsec scales.
Ram Pressure Stripping Models
Analytical and simulation-based study of ram pressure stripping dependence on orbital parameters of galaxies.