Artificial Intelligence Projects in Bangalore

A convolutional neural network fine-tuned on the NIH ChestX-ray14 dataset classifies 14 thoracic pathologies including pneumonia, cardiomegaly, and pleural effusion from frontal chest radiographs. DenseNet-121 is used as the backbone with class-weighted binary cross-entropy loss to handle label imbalance. Grad-CAM visualisations highlight the image regions driving each prediction, providing radiologist-interpretable explanations. The model achieves AUC scores above 0.82 on 11 of 14 pathologies, matching published benchmarks. A web interface allows image upload and returns a ranked pathology report with heatmap overlays.

A deep Q-network agent learns to control traffic signal timings at a four-way intersection simulated in SUMO (Simulation of Urban MObility). The state space encodes queue lengths and waiting times for each lane; the action space selects the next green phase. The DQN agent is trained for 200 episodes and reduces average vehicle waiting time by 34% compared to a fixed-time baseline and 18% compared to an actuated signal controller. The project includes a SUMO visualisation showing the agent's learned policy in real time.

A retrieval-augmented generation chatbot answers student queries about admission procedures, fee structures, and exam schedules using a corpus of 200 college documents. Documents are chunked, embedded with a sentence-transformer model, and stored in a FAISS vector index. At query time, the top-5 relevant chunks are retrieved and passed as context to a locally hosted Mistral-7B model. The system achieves 87% answer accuracy on a 100-question evaluation set. A React chat interface with source citation is deployed as a Next.js application.

An Expectimax search agent plays the 2048 sliding-tile puzzle, modelling the random tile placement as a chance node. The evaluation heuristic combines monotonicity, smoothness, empty-cell count, and maximum-tile position into a weighted score. With a search depth of 4, the agent reaches the 2048 tile in 94% of games and the 4096 tile in 61% of games, compared to 12% and 3% for a random player. The project includes a browser-based game board where users can watch the agent play in real time or compete against it.

A BERT-based regression model is fine-tuned on the Hewlett Foundation Automated Student Assessment Prize dataset to score student essays on a 1–12 scale. The model is trained with mean squared error loss and evaluated using quadratic weighted kappa, achieving a QWK of 0.81 on the validation set, competitive with published baselines. Attention visualisations reveal that the model attends to discourse connectives and topic-relevant vocabulary. A web interface accepts essay text and returns a score with a brief rubric-aligned feedback summary generated by a prompted language model.

A resume screening pipeline parses PDF and DOCX resumes, extracts structured fields (skills, education, experience) using spaCy named-entity recognition, and ranks candidates against a job description using cosine similarity of TF-IDF vectors. A logistic regression classifier trained on 5,000 labelled resume-job pairs predicts shortlisting probability. The system reduces manual screening time by an estimated 70% in a simulated HR workflow. A recruiter dashboard displays ranked candidates with skill-match breakdowns and allows threshold adjustment.

A mobile robot navigation system is implemented in ROS 2 with a simulated TurtleBot3 in Gazebo. A* search on a 2D occupancy grid handles global path planning; a Rapidly-exploring Random Tree planner handles dynamic obstacle avoidance in the local frame. The robot navigates a 10×10 metre warehouse map with 15 randomly placed obstacles, reaching the goal in 100% of trials with an average path length 8% above the theoretical optimum. The project includes a RViz visualisation showing the planned path, explored nodes, and robot trajectory.

A real-time sign language recognition system captures hand landmarks using MediaPipe Hands and classifies sequences of 30 frames into 20 ASL alphabet signs using a stacked LSTM network. The dataset consists of 3,000 self-recorded sequences per sign. The model achieves 96.4% top-1 accuracy on the held-out test set. The system runs at 24 FPS on a standard laptop webcam without a GPU. A live demo overlays the predicted sign and confidence score on the webcam feed, making it suitable for a classroom demonstration.

A fake news detection classifier is trained on the LIAR dataset containing 12,836 human-labelled political statements. RoBERTa sentence embeddings are extracted and fed into a gradient-boosted classifier. The model achieves 74.2% accuracy on the six-class task (true, mostly-true, half-true, barely-true, false, pants-on-fire), outperforming the LSTM baseline by 9 points. A browser extension prototype highlights news article sentences with a credibility score and links to fact-check sources retrieved from a Bing Search API integration.