Hi, the name's Ruhail Rizwan

Impact Statement: Engineered "Little Learner," an interactive AI academic tutor, to provide accessible, voice-based education for Pakistani children (aged 3-7). The system delivers personalized courses and quizzes, overcoming the critical challenge of regional speech variations. The core achievement was developing a custom speech recognition pipeline by fine-tuning the Whisper model on a newly-created, proprietary dataset of Pakistani children's speech.

Key Technologies & Features:

• Core ML Models: Fine-tuned Whisper (Speech Recognition), LLM (Answer Validation), SepFormer (Noise Reduction).
• Core Techniques: Custom Dataset Curation (Pakistani children aged 3-7), Audio Processing (Noise Reduction, Normalization), Intent Classification.
• Application Modules: Voice Interaction & Processing, Content Management & Personalization, Parent/Guardian Progress Reporting.
• Features: Interactive voice-based courses, real-time feedback, personalized learning paths, and a child-friendly web application.

Impact Statement: Developed a robust, production-ready End-to-End MLOps Pipeline for time-series Air Quality Index (AQI) forecasting. The system automates data collection from external APIs, trains a high-performance deep learning model, and implements real-time system monitoring to ensure stability and performance, demonstrating mastery of operationalizing predictive models.

Key Technologies & Metrics:

• Model: LSTM (Deep Learning) for time-series prediction.
• MLOps Stack: DVC (Data Versioning), MLflow (Experiment Tracking), Prometheus & Grafana (Real-Time Monitoring).
• Deployment: Flask API and Docker.
• Performance: Handled 50 requests/second with an average response time of 250ms; achieved Validation Loss of 0.077.

Impact Statement: Engineered a fully dynamic and interactive data visualization dashboard using pure JavaScript and D3.js. The solution allows users to upload custom JSON data and configure four distinct, complex visualizations that are dynamically linked. The core achievement was implementing centralized cross-chart brushing and filtering logic to ensure seamless propagation of user selections across all four charts (Radial Bar, Chord, Force-Directed, Sunburst).

Key Technologies and Metrics:

• Frontend: D3.js, JavaScript, HTML, CSS.
• Core Techniques: Dynamic Data Transformation, Data Binding, Centralized Cross-Chart Interactivity.
• Features: Supports user file uploads (JSON) and custom dimension mapping for immediate visualization rendering.

Impact Statement: Built a cutting-edge Multimodal RAG pipeline capable of processing and answering complex questions based on PDF documents that contain both text and visual information (charts, graphs). The system uses the CLIP model to retrieve contextually relevant chunks (text or image captions) to ground the LLM, effectively reducing hallucinations and demonstrating advanced information retrieval from complex sources.

Key Technologies & Metrics:

• Architecture: Multimodal RAG Pipeline.
• Embedding/Retrieval: CLIP-ViT-B-32 (joint text-image embeddings), ChromaDB (Vector Store).
• LLM: Quantized Mistral-7B-Instruct-v0.1 (4-bit quantization for efficiency).
• Technique: Implemented Chain-of-Thought (CoT) prompting for enhanced reasoning.
• Performance: Achieved a retrieval Hit Rate of ~80% @K=5.

Impact Statement: Successfully implemented the CycleGAN (Cycle-Consistent Adversarial Network) architecture to perform bidirectional, unpaired image-to-image translation between two distinct domains: real person face photographs and hand-drawn sketches. This showcases expertise in complex generative models and the implementation of the cycle-consistency loss to achieve domain adaptation without requiring paired ground truth data.

Key Technologies & Metrics:

• Model: CycleGAN (Generative Adversarial Network).
• Techniques: Adversarial Loss, Cycle Consistency Loss (enforcing A -> B-> A mapping), Unpaired Training.
• Application: Image-to-Image Translation (Face <-> Sketch).

Impact Statement: Conducted an in-depth comparative study by implementing and training a custom Vision Transformer (ViT) from scratch against a deep Convolutional Neural Network (CNN) for image classification. The project provided rigorous quantitative analysis, demonstrating the superior performance of CNNs (achieving ~93.6% accuracy) over ViT (trained from scratch) on small, low-resolution datasets like the CIFAR-10 cat vs. deer subset.

Key Technologies & Metrics:

• Models: Custom CNN (3 Conv Blocks), Custom ViT (4x4 Patch Size, 6 Transformer Layers).
• Framework: TensorFlow/Keras.
• Evaluation: Comparative Analysis, Ablation Study, Accuracy, Loss Curves.
• Insight: Demonstrated understanding of architectural inductive biases (ViT requires massive data; CNN excels with small data).

Impact Statement: A comprehensive portfolio demonstrating foundational expertise across classic and deep learning ML principles. Key achievements include the manual implementation of Rosenblatt's Perceptron from scratch, achieving a high classification accuracy of ~89% on synthetic data, and a custom 2D Convolution operation implemented in NumPy to analyze the theoretical effects of image processing kernels (Sharpen, Blur, Edge Detection).

Key Technologies & Metrics:

• From Scratch: Rosenblatt's Perceptron ( ~89% Accuracy), 2D Convolution (NumPy).
• Models Covered: RNN, CNN, GAN, VAE, Hyperparameter Search (Random Search).
• Core Skill: Deep understanding of algorithm mechanics demonstrated through pure Python/NumPy implementations.

Impact Statement: Implemented and rigorously compared two foundational generative models, a Variational Autoencoder (VAE) and a Generative Adversarial Network (GAN), for image generation (Cats & Dogs). The project featured a significant innovation in the GAN architecture: a custom similarity discriminator that evaluates the relationship between real and generated image pairs to promote fidelity and avoid common generative pitfalls like mode collapse.

Key Technologies & Metrics:

• Models: VAE (Reparameterization Trick, KL Divergence Loss), Custom GAN (DCGAN-like Generator).
• Innovation: Custom Similarity Discriminator Design and Diversity-Promoting Techniques.
• Goal: Generation of higher quality and more diverse images of low-resolution data.

Impact Statement: Developed a Transformer-based Neural Machine Translation (NMT) model from the ground up, implementing all core architectural components—Positional Encoding, Multi-Head Attention, and a dynamic learning rate scheduler—from the original "Attention is All You Need" paper. The system successfully trained on a parallel corpus to translate English sentences to Urdu, with performance rigorously evaluated using the industry-standard BLEU score.

Key Technologies & Metrics:

• Architecture: Transformer (full Encoder-Decoder).
• Core Components: Multi-Head Attention (MHA), Custom Positional Encoding, Custom Learning Rate Scheduler (Warmup Steps).
• NLP Tools: SentencePiece Tokenization (Subword), NLTK.
• Evaluation: Corpus BLEU Score (Observed ~3.55), Masked Accuracy.

Impact Statement: Developed a robust, Dockerized API using FastAPI for resource-efficient, AI-driven recipe search and recommendation. The system uses FAISS vector indexing to enable ultra-fast semantic retrieval of recipes based on ingredients or natural language queries, which then feeds context to a quantized local LLM (Mistral-7B) to generate creative and relevant recommendations, demonstrating a full deployment stack for a generative AI service.

Key Technologies & Metrics:

• Backend: FastAPI, Docker (Containerization).
• AI/NLP: RAG (Retrieval-Augmented Generation), FAISS (Vector Indexing), Quantized Mistral-7B (GGUF).
• Features: Semantic Search, LLM-Powered Generation, User Preference Management.