Hello, I'm Mayank Vyas

Real-time speech + text scam detection on edge hardware, no cloud dependency. Built a multimodal fusion pipeline combining Whisper Tiny (audio transcription) and XGBoost (text classification) with federated learning for on-device model updates.

• Built a multimodal fusion pipeline running at <50ms latency on constrained hardware with zero cloud calls
• Integrated federated learning for on-device model updates — preserving user privacy by ensuring no raw audio ever leaves the device
• Evaluated against adversarial and real-world audio distributions; measured false-positive rate across diverse scam speech patterns

A 4-agent orchestration pipeline using LangChain + LangGraph that autonomously prepares candidates for interviews — parsing resumes, researching companies, generating tailored questions, and building personalized study plans.

• Architected planner/executor agent loops with shared memory and tool routing across 4 specialized agents
• Agents autonomously scrape company data, parse resumes, generate role-specific questions, and build study plans
• Configurable fallbacks and kill-switches for reliable multi-source knowledge routing

Transform educational questions into interactive, story-based visualizations using AI. Features a 4-layer pipeline that intelligently routes content from documents (PDF/DOCX) to 18 distinct game templates with intelligent caching and real-time progress tracking.

• 1st Place Winner at HackASU 2025 (Anthropic Sponsored)
• Built 18 game templates with template-aware story generation
• Implemented intelligent caching reducing processing time by 80%

Research project investigating efficiency, scalability, and linguistic adaptability of Fine-Tuned LLMs for code generation. Explores LoRA rank optimization, data scaling effects, and cross-language generalization using GPT-2.

• Identified optimal LoRA rank 16 achieving 30% syntax pass rate
• Discovered "Complexity Trap" in data scaling behavior
• Demonstrated language-agnostic learning across Python, Java, JavaScript

Python pipeline to convert inspection JSON data into populated TREC (Texas Real Estate Commission) HTML reports. Features smart mapping across 6 TREC sections, automatic empty section removal, and proper formatting for comments, images, and videos.

• Automated mapping of line items to TREC sections I-VI
• Smart filtering removes empty sections automatically
• Proper media embedding with images and video controls

Power BI dashboard with DAX measures and advanced data modeling for actionable business intelligence.

• Achieved 99.8% accuracy in YoY growth calculations
• Reduced query time by 40% through star schema optimization
• Revealed $1.2M revenue opportunity via geo-spatial analysis

Automated data extraction and real-time visualization pipeline for Intel's retail edge computing platform. Built Python scripts to extract metrics from results logs and publish to an MQTT broker, integrated with Grafana dashboards via the MQTT plugin. Created custom Docker images for Grafana and MQTT configured to communicate on the same Docker network using Docker Compose.

• Reduced MTTR by 73% through custom alerting
• Implemented JWT-based OAuth 2.0 with RBAC for SOC2 compliance
• Automated data extraction pipeline with real-time MQTT streaming

This project is part of a Bachelor's Research Thesis, aiming to detect and segment primary roots in plant images using a customized version of the Mask R-CNN model adapted for TensorFlow 2.0 and Keras 2.2.8. The original codebase from Matterport's Mask R-CNN was modified for compatibility and to support training and inference on annotated root datasets.

• Achieved 96.5% IoU accuracy through transfer learning
• Reduced annotation workload by 90%
• Published in Springer's CV in Plant Phenotyping conference

This projects implementation of a Multi-Layer Perceptron (MLP) from scratch using Python. It demonstrates the fundamental concepts of building and training a neural network, including forward propagation, backward propagation, and parameter optimization.

• Achieved 92% accuracy on MNIST using only NumPy
• Implemented automatic differentiation for gradients
• Created interactive weight matrix visualizations