This project introduces an interactive career analytics and resume intelligence system that transforms traditional static resumes into structured, queryable professional profiles. By moving away from "lossy" one-page documents where relevant experience is often fragmented across different sections, the system uses a unified career knowledge model to consolidate education, work history, and projects into a single searchable schema.

Key capabilities include skill-centric analytics that aggregate evidence across all career stages and a question-answering engine that provides employers with concise, evidence-based reports regarding specific skills, such as a candidate's depth of experience with a particular technology. Users and recruiters can interactively "slice and dice" a profile via a web-based dashboard, filtering by technology, role, or time period to generate tailored summaries and eliminate the need for manual resume tailoring. The system leverages LLM-based reasoning grounded in user-provided data to ensure all generated answers are traceable to specific roles and projects, effectively bridging the gap between candidate experience and precise employer inquiries.

The LLMApps Factory is a declarative, low-code platform that simplifies the development of web and mobile applications powered by Large Language Models. By addressing the high engineering effort typically required for LLM integration—such as prompt design, UI wiring, and deploymentthe platform allows students and domain experts to build robust tools without writing custom backend code. Applications are defined using JSON or YAML schemas that strictly separate prompt logic, UI definitions, and deployment concerns, ensuring a cleaner, more accessible development process.

The platform's architecture features a model-agnostic gateway that interfaces with various LLM providers and a dynamic UI renderer that generates interfaces from specified input and output schemas. Supporting multiple formats like Progressive Web Apps (PWAs), mobile wrappers, and embeddable widgets, the system is optimized for single-turn or lightly structured interactions, such as academic assistants or data explainers. This approach prioritizes clarity, reproducibility, and safety through built-in validation and prompt guardrails, providing a scalable foundation for intelligent application design.

The TeacherSim platform is a cutting-edge, generative AI-powered environment designed for deliberate teaching practice through high-fidelity classroom simulations. By leveraging Large Language Models (LLMs) and multi-agent systems, the platform creates a virtual teaching lab where educators can conduct repeatable, goal-oriented sessions without the risks of trial-and-error in real classrooms. The system's architecture comprises a Practice Configurator for setting session objectives, an Interaction Engine for real-time engagement with virtual students, and a Feedback Engine that enables mentor agents to analyze performance. This approach offers a scalable, cost-effective alternative to traditional training and expensive virtual reality systems, enabling the intensive practice required to achieve professional mastery.

The platform features adaptive student agents programmed with diverse learning styles, behavioral traits, and prior knowledge, grounded in established pedagogical frameworks such as Bloom’s Taxonomy and the Big Five personality traits. These agents are not static; they simulate realistic classroom dynamics by exhibiting growth, regression, or spontaneous disruptions in response to a teacher's instructional methods. Complementing these learners are multidimensional mentor agents that provide immediate, structured feedback on critical competencies such as classroom management, inclusivity, and instructional clarity. Through a Performance Dashboard, educators receive data-driven insights and analytics, enabling an iterative cycle of practice and reflection that accelerates the acquisition of complex pedagogical skills.

The Web-to-Visual Narrative Converter is an innovative system designed to address the challenge of consuming long, dense, and cognitively demanding web content. By integrating Large Language Models (LLMs) with diffusion-based image generation, this project provides a service and a Chrome extension that transforms arbitrary web pages into short, engaging visual stories, such as comic-style panels. This tool enables users to quickly grasp the main ideas of news articles, reports, or technical blogs through a concise sequence of images accompanied by short captions that capture the key points.

The system works by extracting the main text, removing boilerplate, and decomposing the information into a structured narrative of 3 to 8 key scenes. A backend pipeline converts these scenes into detailed image prompts, ensuring a consistent visual style—ranging from sketches to infographics—across the panels. The final output is delivered through an intuitive browser interface that lets users easily switch between the original text and its visual summary, prioritizing content fidelity and narrative coherence.

The LLM-Driven Autonomous Analytics Dashboard Generation from SQL Databases project is designed to bridge the gap between complex relational data and actionable business insights by automating analytical reasoning. Traditional data extraction is often *time-consuming and highly dependent on skilled analysts who can write SQL and design visualizations; furthermore, many business users may not know which questions are most useful to ask of their data. This project leverages Large Language Models (LLMs) to independently infer user intent and organizational goals, enabling the system to determine which metrics to track, which queries to execute, and how best to visualize results without explicit, step-by-step instructions from the user.

To achieve this, the system uses a modular pipeline that starts by analyzing a database schema to identify entities and relationships. It then interprets high-level natural-language descriptions of the user's role, business domain, and organizational objectives to automatically propose relevant KPIs and trends. The architecture includes specialized components for safe SQL query generation, automated visualization selection, and dashboard composition, ensuring the final output is a coherent, role-aware interface. By moving beyond simple text-to-SQL translation, the project aims to lower the barrier for non-technical users and significantly reduce the manual effort required for exploratory data analysis.

This project involves the design and implementation of an LLM-assisted platform to create expert-driven interactive decision wizards. Traditional methods for encoding expert judgment, such as informal checklists or rigid rule engines, are often brittle, difficult to maintain, or lack the transparency required for safety-critical workflows. To address these limitations, the sources describe a hybrid system that combines expert-defined logic with Large Language Models (LLMs) used exclusively as constrained decision operators. The platform is divided into two main components: an Expert Wizard Builder, where domain experts define decision steps without writing code, and a User Wizard Runner, which guides end users through a structured, point-and-click interface to achieve a specific outcome.

The core functionality of the system relies on LLMs to interpret ambiguous input or perform bounded classification according to strict schemas, rather than allowing them to act as free-form conversational agents. Crucially, the LLM is prohibited from generating new steps, modifying routing logic, or providing unbounded explanations, ensuring that all decision outcomes remain explainable, deterministic, and traceable to specific inputs. The architecture includes a web-based frontend and a backend execution engine that manages different step types, such as Collect, Decide, and Outcome. By safely integrating LLMs into decision-support workflows, the project demonstrates a practical approach to software engineering and the responsible use of AI in domains such as medical intake, HR screening, and technical troubleshooting.

Music-Driven Video Augmentation aims to bridge signal processing and computer vision by synchronizing low-latency audio analysis with lightweight video effects to enable immersive AR experiences. Unlike static or manual editing, this system extracts musical features, such as BPM, onset timing, and spectral energy, to drive real-time visual transformations. A critical requirement is maintaining user safety and scene visibility, ensuring that the augmentation, such as edge amplification or contrast modulation, subtly enhances the stream without obscuring the environment or causing visual fatigue from excessive flicker.

The system architecture comprises four distinct modules: Audio Analysis, Control and Mapping, Video Augmentation, and a Real-Time Execution Layer. These modules work together to convert audio triggers into frame-local video effects, including Sobel filtering, gamma modulation, and motion-aware gating while maintaining a strict latency target of 50–100 ms. By avoiding resource-intensive deep learning models in favor of classical image processing and simple shaders, the project ensures high performance and deterministic behavior on commodity hardware or AR-class devices. Success is measured by evaluating the synchronization accuracy, system stability, and qualitative perceptual quality to provide a comfortable and responsive user experience.

This project proposes a web-based interactive system that simplifies high-quality image editing by integrating object detection and diffusion models. While traditional generative tools often require users to manually draw complex masks a task that is both difficult for non-experts and error-prone this application automates the process by identifying objects and providing pixel-level masks or bounding boxes for selection. By combining these technologies, the system enables an intuitive workflow in which users perform object-level edits by selecting detected regions and providing text prompts to guide the generative inpainting process.

The system architecture is divided into a React-based frontend for image visualization and a Python backend (using FastAPI or Flask) that manages AI model inference and mask processing. The functional pipeline involves uploading an image, automatically generating object labels and confidence scores, and applying Stable Diffusion-based inpainting to the user’s chosen regions. This approach yields a modular and reliable tool that enables users to compare the original and edited versions before downloading the final result, demonstrating a practical integration of computer vision and generative AI.

This project proposes a "Crowd-Sourced Generative Feed Platform" designed to transform AI usage from a tool for isolated individual productivity into a medium for collective creativity. Currently, most AI prompts and insights are executed once and discarded; this platform instead treats them as reusable building blocks, akin to code in an open-source ecosystem. By enabling users to publish "generative feeds" defined as combinations of input data, prompts, and transformation rules, the system creates "living artifacts" that allow others to build creatively on existing ideas rather than starting from scratch.

The platform functions as a shared creative substrate, visualizing the evolution of these feeds as a directed graph of creative transformations. This structure supports advanced use cases such as collective news sense-making, where users can add layered thematic framing to raw data, or visual idea amplification, which bridges textual and visual reasoning through iterative design. By ensuring that updates in source feeds propagate automatically to derived feeds, the project creates a dynamic environment where creative logic is explicit, composable, and continuously evolving through community collaboration.

BrowseAndLearn revolutionizes digital language acquisition by embedding an LLM-powered browser extension directly into your everyday web surfing, eliminating the need to separate learning from authentic content like news or documentation. Unlike traditional tools that rely on full translations, this system preserves the natural reading flow by offering selective, context-aware assistance only when necessary. By identifying specific linguistic hurdles, such as polysemous words, idioms, and complex grammar, and providing minimal guidance, the project aims to enhance long-term retention and motivation while minimizing cognitive load.

The core innovation lies in its use of Large Language Models (LLMs) to act as a dynamic, real-time tutor that adapts to your individual proficiency. The system analyzes web text to generate personalized interventions, ranging from short mother-tongue glosses to dynamic inline micro-assessments such as fill-in-the-blank exercises. Through a continuous feedback loop, the AI interprets your interactions to update its model of your knowledge, automatically fading assistance for mastered vocabulary while deepening explanations for persistent challenges

This project tackles the liquidity limitations inherent in traditional direct barter systems, where the low probability of finding a perfect mutual match often leaves significant functional value trapped in underutilized assets. By implementing an AI-driven swapping platform, the system unlocks this value through multi-party exchange chains, allowing items to move through a directed network of users to reach those who desire them most, rather than relying on restrictive pairwise trades. This approach dramatically increases the volume of successful exchanges, ensuring that users can effectively liquidate items they no longer need in exchange for goods that provide greater personal value.

To optimize user satisfaction, the platform employs semantic analysis and predictive modeling to infer latent preferences and estimate the specific utility of potential items for each participant. The core algorithmic engine constructs a directed swap graph weighted by these utility scores and applies Integer Linear Programming and cycle enumeration techniques to maximize the total predicted utility and the number of completed swaps. Furthermore, the system integrates a trust and reliability model to predict dropout risks and penalize fragile chains, ensuring efficient execution and high confidence in the value transfer process

This project tackles the liquidity limitations inherent in traditional direct barter systems, where the low probability of finding a perfect mutual match often leaves significant functional value trapped in underutilized assets. By implementing an AI-driven swapping platform, the system unlocks this value through multi-party exchange chains, allowing items to move through a directed network of users to reach those who desire them most, rather than relying on restrictive pairwise trades. This approach dramatically increases the volume of successful exchanges, ensuring that users can effectively liquidate items they no longer need in exchange for goods that provide greater personal value.

To optimize user satisfaction, the platform employs semantic analysis and predictive modeling to infer latent preferences and estimate the specific utility of potential items for each participant. The core algorithmic engine constructs a directed swap graph weighted by these utility scores and applies Integer Linear Programming and cycle enumeration techniques to maximize the total predicted utility and the number of completed swaps. Furthermore, the system integrates a trust and reliability model to predict dropout risks and penalize fragile chains, ensuring efficient execution and high confidence in the value transfer process

This project addresses the challenge of consuming long, dense technical documents such as textbooks, research papers, and regulatory texts by automatically transforming static PDFs and web pages into a structured, wiki-like knowledge space. Rather than relying on linear reading, which creates a high cognitive load when concepts are scattered across sections, this system facilitates "study by navigation and reasoning." It reorganizes content so that users can view central concepts as interconnected nodes, allowing them to revisit definitions, connect ideas across multiple documents, and read high-level summaries before diving into specific details.

To achieve this, the system employs Large Language Models (LLMs) to detect topic boundaries, infer logical hierarchies, and generate summaries at multiple levels, ranging from corpus overviews to section-specific explanations. The methodology goes beyond simple formatting by leveraging semantic understanding to link key technical terms and abstract ideas, and by performing cross-document reasoning to identify overlaps, differences, or contradictions across sources. The final result is an interactive interface that reduces the effort required to reconstruct context, enabling students to progressively expand their understanding through a graph of semantic links and expandable summaries