Things I’ve Built
Every project was built to explore a real problem, test an engineering hypothesis, and learn from physical and digital failure modes.
MessOS
A software system designed to make hostel mess operations simpler, more transparent and easier to manage.
Hostel mess operations suffer from manual attendance tracking, untracked dietary wastes, chaotic meal complaints, and non-transparent billing across managers, munshis, and students.
A unified B2B SaaS platform including Mess Manager, Munshi, and Clerk portals for real-time meal counts, transparent accounting, feedback loops, and automated attendance.
Kavach
AI-powered voice authentication and physical smart access system.
People frequently forget physical keys, while traditional contact biometrics require physical touch and specialized optical sensors. We asked: What if human voice could securely become the key?
Turn human voice into a biometric key by integrating real-time deep learning speaker verification with a Raspberry Pi server, Arduino microcontroller, and electromechanical motor actuators.
NanoTrade
Real-time paper trading platform with custom matching engine.
Most trading demos are simple database dashboards that fail to simulate realistic order-book mechanics, execution latency, queue matching, and portfolio state reconciliation.
Understand what happens underneath a real trading interface by building an event-driven paper trading platform powered by a custom matching engine.
SkillGap AI
Semantic embedding and RAG engine for resume and job description fit analysis.
A résumé and a job description can superficially look like a match while still concealing critical capability gaps, leading to poor hiring signals and unfocused career preparation.
Use semantic vector embeddings and Retrieval-Augmented Generation (RAG) to measure genuine competency fit, detect latent skill gaps, and generate structured learning roadmaps.
MovieSentiment
Production-oriented ML discovery platform with automated MLOps & EKS deployment.
Machine learning models often remain trapped in Jupyter notebooks without robust CI/CD, containerized packaging, automated retraining, and scalable Kubernetes orchestration.
Build an end-to-end ML sentiment discovery platform demonstrating the full lifecycle: Model → Version → Package → Deploy → Monitor.
Bike Demand Prediction ML System
Production ML demand forecasting application developed at TS Bridge.
Urban mobility fleets experience extreme utilization volatility due to weather, seasonal shifts, and commuting patterns, requiring precise predictive demand modeling to prevent asset shortages.
Train, evaluate, and deploy an end-to-end regression pipeline on AWS with automated MLOps tracking and stakeholder analytics dashboards.
Large-Scale Data Preprocessing & EDA
Data wrangling and feature engineering pipeline on 50,000+ records at CourseVita.
Raw real-world datasets arrive laden with missing values, inconsistent encodings, extreme outliers, and duplicate records that degrade downstream modeling.
Construct a robust, reproducible data wrangling and exploratory analysis pipeline to transform 50,000+ messy records into clean, modeling-ready datasets.
Vehicle Insurance ML Pipeline
End-to-end predictive classification pipeline with modular MLOps architecture.
Assessing vehicle insurance claim propensities requires handling severe class imbalances, multi-modal features, and repeatable retraining pipelines.
Build a modular, reproducible ML pipeline implementing structured feature engineering, hyperparameter tuning, and containerized artifact tracking.
Industrial Sensor Fault Detection
Anomaly detection system for industrial sensor telemetries.
Industrial machinery sensor drifts and electrical spikes often go undetected until catastrophic mechanical failure occurs.
Apply time-series statistical anomaly detection and threshold heuristics to continuous sensor telemetry streams.