Saisrijith Reddy — AI • ML • Data Science Project Portfolio
🚀 Featured Projects
Below is a curated lineup of my most impactful AI, ML, and data science projects. Each link takes you directly to the project’s GitHub repository or article.
🧠 MIRA — Multimodal AI Assistant
GitHub: Private Build
Medium Article: Building MIRA
A fully modular, voice-activated AI assistant with many autonomous agents (BrowserAgent, PlannerAgent, GmailAgent, CalendarAgent, WhatsAppAgent, MusicAgent).
Includes wake-word detection, GPT-4o vision fusion, trust-ranked browsing, async orchestration, and multimodal reasoning.
🏠 Housing Price Prediction (1st Place — University Competition)
Website: Housing Price Prediction
600k+ record real-estate dataset.
Feature engineering + binning + Optuna-optimized XGBoost/LightGBM.
🥊 UFC Fight Prediction — Leakage-Free ML Pipeline & Inference App
GitHub: UFC Fight Prediction
Website: Live App
An end-to-end, leakage-free UFC fight prediction system built under strict prefight constraints, covering data curation, feature engineering, model training, calibration, and deployment.
The pipeline enforces temporal integrity using Glicko ratings, shift-based feature construction, quantile clipping, and log transforms, ensuring all features are legally available at prediction time.
Multiple models are evaluated, with a calibrated XGBoost model achieving strong out-of-sample AUC and well-behaved probabilities, making it suitable for real-world inference.
The project culminates in a production-safe Streamlit app that mirrors training preprocessing exactly and outputs deterministic, calibrated win probabilities for full fight cards.
🧠 Sentiment Classification with Neural Networks
GitHub: Sentiment Classification
CNN-based sentiment classification on the Sentiment140 dataset (1.6M tweets).
Baseline CNN outperformed deeper variants; best test accuracy (79.54%) achieved with a 500-dimensional embedding, showing hyperparameters matter more than model depth.
🎬 CineSeq — Decay-Aware Seq2Seq Forecasting
GitHub: CineSeq
Research-oriented framework for short-horizon movie box-office forecasting.
Combines explicit exponential decay modeling with multivariate external signals using an LSTM–GRU Seq2Seq architecture.
Builds a leakage-safe weekly panel (2010–2023), validates decay dynamics, and forecasts 4-week revenue horizons.
Incorporates trailer-based emotion embeddings via transfer learning and interprets predictions using Integrated Gradients.
Attention-based decoding reduces test WAPE by ~41%, with strongest gains for movies exhibiting mid-run deviations from smooth decay.
🎾 Elo-Based Time Series Forecasting
GitHub: Elo Forecasting Repo
ATP tennis forecasting dashboard using ARIMA auto-selection + surface-aware Elo system.
Forecasts 6–24 month player form and win probabilities.
🏀 Fair NBA Draft Predictor — FairStacks Ensemble
Website: FairNBA Draft Predictor
Fairness-aware classification using Naive Bayes, LDA, SVM, Ridge/Lasso Logistic Regression with CVX-based fairness constraints.
Reduces TPR gaps across school tiers.
📉 Bias–Variance Tradeoff in Ridge Regression
Website: Bias-Variance Tradeoff
Simulated Ridge Regression vs OLS under linear and nonlinear DGPs.
Demonstrated full bias–variance decomposition with clear visualizations showing when Ridge outperforms OLS.
💳 Payment Method Classification (ML + ANN)
GitHub: Payment Method Classification
Leakage-free classification of payment method (Credit Card, Debit Card, PayPal) using numerical retail transaction features.
Multinomial logistic regression outperformed an ANN (≈0.67 vs ≈0.63 test accuracy), indicating limited signal rather than model instability.
📊 Global Safety Regression Model
GitHub: Global Safety Regression Repo
Analyzing how GDP, internet access, population growth, refugees, and development metrics relate to homicide rates across 169 countries.
📏 Multivariate Body Fat & BMI Analysis
GitHub: Multivariate Stats Repo
Multivariate regression, MANOVA, PCA, and CCA on physical measurements to estimate body fat % and BMI.
📊 End-to-End Quant Alpha Pipeline (Layered Workflow)
Website: Quant Alpha Pipeline
Developed as a summer research project, this full-stack workflow builds and evaluates equity alpha factors across multiple time horizons.
Integrated OHLCV, options, and macroeconomic data, and applied SHAP- and IC-based feature pruning, ElasticNet ensembles, and Optuna-tuned XGBoost/LightGBM models to forecast 1D/5D/21D forward returns.
Extended with HMM-based regime detection and CVXPY portfolio optimization, achieving robust out-of-sample IC and Sharpe improvements.
💼 Microsoft × Pinterest Valuation (Imperial College Project)
GitHub: Valuation Analysis Repo
Full DCF, comps, and precedent analysis determining an implied acquisition price of $26.87/share.
Includes a 30-page report and pitch deck.
📬 Contact
GitHub: github.com/srijith-reddy
LinkedIn: linkedin.com/in/saisrijith-reddy-maramreddy-399869166