Computer Science, ISE, AI & Machine Learning in Bangalore

A computer vision system that identifies registered individuals from a live camera feed and marks attendance automatically in a database, eliminating manual roll-calls. Uses MTCNN for face detection and FaceNet embeddings for identification, achieving sub-200ms recognition latency on standard hardware.

A CNN-based image classification system trained on PlantVillage dataset to identify 38 crop disease categories from leaf photographs captured via a mobile camera. The system runs inference on-device using TensorFlow Lite, enabling offline diagnosis in low-connectivity agricultural settings.

An NLP pipeline that classifies news articles as real or fabricated using a fine-tuned BERT model trained on the LIAR and FakeNewsNet datasets. A browser extension and web dashboard allow users to paste article text or URLs for real-time credibility scoring with explainability outputs.

A decentralised land ownership and transfer system built on Ethereum, where property records are stored as NFTs and ownership transfers are executed via smart contracts — removing the need for centralised intermediaries and reducing fraud risk. Includes a React frontend connected to MetaMask.

An automated recruitment pipeline that parses resumes in PDF/DOCX format, extracts structured entities (skills, experience, education) using spaCy NER, and ranks candidates against a job description using TF-IDF and semantic similarity scoring. Reduces initial screening time by over 80% in test deployments.

A real-time alert system that monitors a driver's eye aspect ratio (EAR) and mouth aspect ratio (MAR) using dlib facial landmarks to detect microsleep and yawning events. An audio alarm triggers when drowsiness thresholds are exceeded for more than 2 consecutive seconds, with an optional SMS alert via Twilio.

A privacy-preserving machine learning framework where multiple hospital nodes train a shared diagnostic model on local patient data without centralising records. Uses PySyft for federated aggregation and demonstrates differential privacy guarantees on a chest X-ray pneumonia classification task.

A time-series forecasting system that predicts urban traffic density at intersections 30 minutes ahead using LSTM networks trained on historical sensor data. The model integrates with a simulated traffic light controller to demonstrate adaptive signal timing, reducing average wait times by 22% in simulation.

A retrieval-augmented generation chatbot that ingests institutional documents (syllabi, regulations, timetables) into a vector database and answers student queries with cited source passages. Built on LangChain with a local LLM backend, ensuring no student data is sent to external APIs.

A binary and multi-class classifier trained on the NSL-KDD and CICIDS datasets to distinguish normal network traffic from 14 attack categories including DoS, probe, R2L and U2R. An XGBoost ensemble achieves 98.4% accuracy on the test set, with a live packet capture interface using Scapy.

An automated warehouse inventory system using RFID readers and ESP32 nodes to track item movements in real time. A web dashboard displays current stock levels, triggers low-stock alerts via Telegram, and logs all transactions to a cloud database — eliminating manual stocktaking for small warehouses.

A real-time pipeline that streams tweets via the Twitter API, classifies sentiment using a fine-tuned RoBERTa model, and visualises trends on a live dashboard. Supports topic filtering, temporal trend plots, and geographic heat maps of sentiment distribution.