Charvi Kusuma

Magnecruit: AI-Powered Productive Recruitment Workspace Application

Architected and built a comprehensive recruitment tool featuring an AI assistant (MagnecAI) to enhance recruiter productivity.

AI & Backend: Developed the backend using Python/Flask, integrating Google Gemini API for natural language interaction. Implemented agentic logic to orchestrate LLM calls, parse responses (JSON extraction via regex/parsing), manage conversation state, and interact with the PostgreSQL database via SQLAlchemy ORM. Utilized Flask-SocketIO for real-time bidirectional communication. Designed RESTful endpoints for user authentication and conversation management.

Frontend: Created a responsive user interface with React and TypeScript, featuring distinct panels for navigation, AI chat, and a dynamic workspace. Implemented real-time updates in the workspace component using Socket.IO, allowing users to see AI-generated sequences appear live. Managed component state effectively using React Hooks.

Key Skills: Python, Flask, Flask-SocketIO, SQLAlchemy, Google Gemini API, Prompt Engineering, JSON Parsing, React, TypeScript, Socket.IO Client, REST APIs, PostgreSQL, Git, HTML, CSS (Tailwind).

Resonique - Multimodal Music Recommendation

Integrated Gemini Flash to semantically describe user surroundings (text/image/audio) and Spotify song metadata, transforming them into high-fidelity vector embeddings using MPNet for emotion-based music retrieval.

It leverages

1. Pinecone for vector similarity search
2. Google Gemini AI for understanding user queries
3. Supabase for cloud audio storage
4. Spotify API for real-world song recommendations

Key Skills: Multi-modal Search, Generative AI, LLM, Vector Search, MPNet, CLAP, Pinecone, Supabase, API Integration, Streamlit.

Adaptive Driver Assistance: Context-based Approach to Pedestrian Safety

Ongoing research publication focuses on utilizing Parameter Efficient Fine-Tuning on ViT Transformer for Pedestrian Behavior & Scene Context Classification

Achieved a significant reduction in trainable parameters to 0.68% of the Vision Transformer (ViT) backbone while maintaining an impressive 90% classification accuracy across four intention adapters.

1. Processed 346 HD video clips from the JAAD (Joint Attention for Autonomous Driving) Dataset, employing YOLOv8 as the pedestrian detector to extract Regions of Interest (ROIs). This approach ensures precise identification of pedestrians in various scenarios, enhancing the model's ability to predict intentions accurately.
2. Implemented a robust pipeline for data preprocessing and model training, including data augmentation techniques to improve generalization and effective performance with Low Rank Adaptation (LoRA) for fine-tuning.
3. Deep Learning: Optimized Vision Transformer, Object Detection, YOLOv8, Deployable Adapters

F.E.A.S.T - Food & Ingredient AI Suggestion Technology

With an aim to reduce food waste and provide optimized meal planning, we performed extensive data augmentation and custon object detection for Ingredients images.

Tokenized 2.2 million recipes for high-quality model inputs for personalized recipe generation and nutritional value provision.

Showcased Project in CSE Demo Day Spring 2024 as "Eyes on Eats: From Image to Formula".

1. Large Language Models, YOLO models (versions 7 and 9), Grounding DINO, BART tokenizer
2. Trained BART transformer model and implemented ChefTransformerT5 and BertFDA-Nutrition-ner model from Hugging Face.
3. Deep Learning: Object Detection, Text Generation, Named Entity Recognition

RxRovers - Roaming for Rapid Relief

Path Optimization and Dynamic Obstacle avoidance strategy simulating efficient delivery of medicines.

Implemented 6 Deep RL algorithms including Value-based and Actor Critic with comphrehensive comparison.

1. Q Learning, Deep Q Networks, A2C, Proximal Policy Optimization (PPO) and improved versions.
2. Reinforcement Learning, Gymnasium Library and Python
3. Beginner's Repo : Defining and Solving RL Environments for Stock Trading.

Immigration Reforms Sentiment Analysis with SenticNet APIs

Worked with 25K tweets for concept parsing, polarity sentiment and word embeddings with Gensim's Word2Vec models to analyze similarities.

Compared multiple NLP tools including TextBlob, VADER, and SenticNet APIs for detailed error analysis of polarity classification and concept parsing.

Implemented Semantic Similarity Analysis (SSA) using unsupervised learning to predict sentiments.

1. Natural Language Processing: TextBlob and VADER, Gensim, Scikit-learn, Tweepy
2. This Project was successfully completed as part of my internship at NTU Research under the guidance of Prof. Erik Cambria

Adaptive Driver Assistance: Context-based Approach to Pedestrian Safety

Applied task-specific tuning to retain base model knowledge, training 8 adapters that reduced parameters to 0.68% of the Vision Transformer backbone while maintaining 90% accuracy in context and behavior classification.

Mapping Crime Dynamics: Integrating Textual, Spatial, and Temporal Perspectives

Developed classification models using algorithms like bagging, boosting, Random Forest, and XGBoost to predict crime types based on various features. Achieved classification accuracies up to 73%, offering a strong foundation for identifying high-risk areas.

Leveraged LSTM (Long Short-Term Memory) networks for time-series analysis, successfully predicting future crime trends by analyzing historical crime data.

Conducted topic modeling using LDA on crime descriptions to uncover latent crime themes, followed by applying clustering algorithms like BIRCH and Mini Batch KMeans to spatially analyze these topics.

1. Machine Learning, LSTM, Topic Modelling

Automated Monitoring System for Healthier Aquaculture Farming

Achieved 88% dead fish detection accuracy using UAVs and DL models for real-time monitoring of aquaculture facilities.

Addressed challenges of inclement weather and large pond coverage using drones equipped with night vision cameras.

1. YOLOv5 variants, Alerting Mechanism, IoT Devices
2. UAV surveillance system, tested in actual shrimp and fish farming facilities.

Attention based Discrimination of Mycoplasma Pneumonia

Implemented attention-based feature extraction to enhance classification of pneumonia, achieving high-dimensional feature entanglement.

Unsupervised Generative Transfusion Network (UGTN) and transformers with 8-layer encoders and decoders.

1. TensorFlow, PyTorch, OpenCV
2. Keras, Python
3. Dataset: COVID-19 Radiography Database

Snake Detector and Alerting Gadget for Rural India Using Yolo

Deployed an Autonomous Snake Detection Device capable of real-time visual recognition for rural India.

Demonstrated expertise in integrating IoT devices with advanced object detection frameworks to design a cost-effective, reliable solution for use in low-light conditions.

1. YOLOv5 variants, Alerting Mechanism
2. Data Augmentation, Object Detection, IoT Integration

Python Based Motion Sensing Digital Writing Pad

Developed a cost-effective, non-touchscreen digital writing pad using motion-sensing techniques, aimed at providing accessible technology to students and educators in economically deprived regions.

Autocorrect, spell check, customizable pen colors, eraser, and save options in JPEG or PDF format

1. Python, OpenCV