A Transformer–RAG framework for SMS scam detection, eight-class scam categorisation, phishing URL scoring, and explainable safety guidance for users in India.
96.4%accuracy
91.2%macro-F1, 8 scam classes
97.8%phishing-URL F1
Transformers
RAG
PyTorch
NLP
The problem
SMS scams in India change wording faster than keyword filters can keep up. Flagging a message is only half the job — people also need to know what kind of scam it is and what to do next.
Approach
A dual-head shared-encoder Transformer classifies each message as scam or legitimate and assigns one of eight scam categories in a single pass.
Links inside the message are scored separately for phishing risk.
A retrieval-augmented advisory layer pulls relevant safety guidance and explains why the message was flagged, in plain language.
Results
96.4% accuracy, 91.2% macro-F1 across the eight categories, and 97.8% F1 on phishing URL detection.
Source · coming soonWith Prof. Sanchit Tyagi
groundwater-forecasting
Time series · Climate · 2025 · Under review
Groundwater forecasting, Italy
Time-series analysis of Italian groundwater levels, comparing classical, ML and deep-learning forecasters for groundwater management and climate-risk analysis.
5models compared
RMSE · MAEevaluation
ARIMA
Prophet
LSTM
XGBoost
GARCH
The problem
Groundwater levels drive irrigation, drinking-water and drought planning. Picking the wrong forecaster means planning around the wrong future.
Approach
Analysed groundwater time-series data and implemented ARIMA, Prophet, LSTM, XGBoost and GARCH forecasters.
Compared every model on RMSE and MAE to find which approaches hold up for management and climate-risk use.
Source · coming soonWith Dr. Manoj Kumar
flappybird-rl
Reinforcement learning · 2026
FlappyBird RL
A Deep Q-Network that learns when to flap — trained with experience replay inside a Flappy Bird environment I built in Pygame.
8state features
2actions: flap or don’t
Python
PyTorch
Pygame
NumPy
DQN
The agent
Deep Q-Network with an 8-feature state representation and a 2-action space.
Experience replay, mini-batch training and reward shaping, implemented in PyTorch.
The environment
Built the game itself in Pygame: physics, pipe generation, collision detection and scoring — so the agent and the world could be tuned together.
Reward design
Positive reward for staying alive, passing pipes and moving toward the gap centre; negative reward for hitting pipes or the screen edges.
Training scores are plotted live with Matplotlib to watch the agent improve.
A voice-controlled Python automation toolkit for weather, application, Bluetooth and file management on macOS. Built with Vedant Singh Bedi and Harsh Modi.
Python
SpeechRecognition
pyttsx3
Gemini
WeatherAPI
AppleScript
What it does
Listens with speech recognition, replies with text-to-speech, and runs system commands without touching the keyboard.
Reports real-time weather, rain chance and AQI through WeatherAPI, auto-detecting location via macOS CoreLocation.
Opens and closes apps, toggles Bluetooth and creates, renames, moves or deletes files — all by voice, through AppleScript and Python.
Uses Google’s Gemini API for open-ended questions.
Played the data scientist at a Boston housing agency: explored U.S. Census housing data, visualised it, and used hypothesis tests and regression to answer management’s questions.
p < 0.001river homes priced higher
r = 0.76NOX vs industry
R² = 0.06distance vs value
Python
Pandas
NumPy
SciPy
statsmodels
Seaborn
What I tested
T-test: homes bounding the Charles River have a significantly different median value (t = −4.0, p ≈ 7×10⁻⁵).
ANOVA: median value differs across age groups of the housing stock (F = 36.4, p ≈ 2×10⁻¹⁵).
Pearson: nitric-oxide levels rise with industrial land share (r = 0.76).
OLS regression: distance to employment centres explains little of home value on its own (R² = 0.06).
