Engineers working with relational databases quickly learn that the main difficulty is not SQL syntax but understanding the hidden logic of a specific schema: implicit relationships, nullable fields, constraints, and unexpected cardinalities that only surface during debugging. Conventional text-to-SQL systems ignore this learning process, treating each query as an isolated generation task and failing to retain experience from prior execution errors.

LLM-SQL-Mind turns text-to-SQL into a continual learning cycle. An LLM Query Generator is coupled with a SQL Executor & Validator that runs queries, captures structured feedback from error messages and results, and extracts latent schema knowledge such as join patterns and filtering rules. These insights are stored in persistent vector memory and retrieved via RAG to ground future prompts, enabling the system to accumulate database-specific expertise over time and progressively improve execution accuracy.

Keywords: LLM, Vibe-Coding, Text2SQL, Agents

Small autonomous drones pose a serious threat to dismounted soldiers, who cannot carry heavy radar-based detection systems. This project proposes a lightweight alternative: a wearable sensing framework called the Dynamic-Array Acoustic World Model (DAWM). Tiny microphones embedded across a soldier’s uniform capture sound from multiple positions. Because the sensors constantly move with the body, the system uses generative AI world-model techniques to adapt to changing sensor geometry and motion. It learns a compact representation of the acoustic scene and predicts evolving Acoustic Energy Maps, allowing it to track and localize approaching drones even under movement and battlefield noise.

Since no real dataset exists for such dynamic, body-mounted sensor arrays, the project relies on simulation-based data generation. Human motion simulators (e.g., OpenSim) are combined with acoustic simulation tools (e.g., SonicSim) to produce realistic audio streams that reflect limb motion, body orientation, and environmental reverberation. These large-scale synthetic datasets are used to train and validate the DAWM model, demonstrating the feasibility of continuous, short-range drone detection using low-power wearable sensors.

Keywords: Signals, Acoustics, Time-Series, Generative AI and World Models, Defense AI

Micro-learning enables flexible, bite-sized learning, but learners often lack the background knowledge needed to fully understand each module. Educators must manually identify and link prerequisite materials, a time-consuming and unscalable process that limits coherence across courses. This project addresses that need by introducing an automated Prerequisite Information Retrieval (PIR) system. Using Large Language Models (LLMs), the system detects hidden conceptual dependencies within texts and automatically generates contextual hyperlinks to the most relevant foundational resources.

Unlike traditional prerequisite learning methods, which classify binary concept relationships, PIR retrieves complete documents that provide the background needed for comprehension. The project establishes a domain-agnostic benchmark dataset through synthetic data generation and structured reuse of Wikipedia content, and evaluates LLM-based retrieval models that combine concept extraction with embedding-based similarity matching. The result is a scalable framework that reduces educator workload while improving conceptual continuity and depth in micro-learning environments.

Keywords: LLM, Education, Text Analysis

Cyber-Physical Systems (CPS) combine physical processes with digital monitoring in safety-critical areas such as predictive maintenance and healthcare. Detecting anomalies in these systems is difficult because they operate across multiple regimes, produce complex sensor data, and rarely provide real failure examples for training. Traditional evaluation methods rely on average performance metrics, which do not reflect worst-case risks or regime shifts, making deployment in mission-critical settings unreliable.

This project proposes a latent-space evaluation framework based on generative models such as Variational Autoencoders (VAEs). Multivariate time-series data are mapped into a latent space that captures different operating regimes. The framework then generates realistic synthetic anomalies via controlled latent manipulation, uses held-out regimes as stress tests, and evaluates models within each latent cluster rather than globally. This enables more reliable model selection and risk estimation even when real anomaly data are scarce

Keywords: Anomaly Detection, Time Series, Signals, Predictive Maintenance, VAE, Latent Representation, Generative AI

Large Language Models can generate code, but most systems validate correctness only after producing an entire program. This generate-then-test approach allows small semantic mistakes to accumulate, resulting in code that appears correct but fails at runtime.

This project proposes Statement-Transactional Execution (STE), an execution-first framework for program synthesis. Instead of generating full functions at once, the model produces one statement at a time, executes it in a sandboxed environment, and keeps it only if it leads to a valid state change. Invalid steps are rolled back, preventing error propagation. By combining controlled execution, state tracking, and an LLM-based validator, STE ensures that code correctness is grounded in observable runtime behavior rather than post hoc testing.

Keywords: LLM, Vibe-Coding, Agents

This project develops an AI system that transforms Zoom video recordings into structured behavioral insights. It detects features such as gaze direction, attention level, phone or headphone use, object interaction (e.g., pens or notebooks), environmental conditions, and the presence of others.

A key component is a synthetic data pipeline built with Stable Diffusion to generate realistic training scenarios that are difficult to collect in real classrooms. The system integrates computer vision models, gaze estimation, pose detection, and multimodal analysis. The result is an end-to-end behavioral analytics framework that produces interpretable reports while demonstrating how synthetic data improves robustness when real labeled data is limited.

Keywords: Computer Vision, Stable Diffusion, Generative AI, Education, Gaze, Object Recognition, Image Classification

Finding a specific person in a dense crowd using only a reference image is a difficult and operationally important vision problem. Current person re-identification methods work best in structured settings with limited occlusion and fail in crowded scenes where individuals are small, partially hidden, or visually similar.

Progress is limited by the lack of scalable training data, as collecting real-world crowd-search datasets poses privacy and logistical challenges. This project proposes a synthetic data engine that inserts a target identity into crowded images using identity-preserving diffusion inpainting. The generated training pairs enable learning a query-conditioned model that locates a person in a scene from a single reference image, testing whether controlled synthetic identity injection can support real-world generalization.

Keywords: Computer Vision, Stable Diffusion, Object Detection, Defense AI, Multimedia Search, Generative AI

Accurate Parkinson’s disease (PD) severity scoring increasingly relies on pose-based analysis of patient movement. However, most existing methods assume reliable skeleton extraction under controlled recording conditions. In real homes, this assumption collapses: cameras are mounted at arbitrary angles, lighting is inconsistent, furniture causes occlusion, and videos are low-resolution or motion-blurred. Under such conditions, pose-only pipelines become unstable and clinically unreliable.

This project proposes a synthetic-data-driven framework for robust regression of Parkinsonian movement severity directly from RGB video under realistic home-camera distortions. We will generate a large paired dataset using neuromotor motion simulation with controllable symptom injection (tremor, bradykinesia, freezing of gait). Each simulated pose sequence will be rendered into photorealistic home environments using pose-conditioned image generation (e.g., ControlNet-Pose) with aggressive domain randomization: occlusion by furniture, unusual viewpoints, motion blur, compression artifacts, and lighting variability. Crucially, ground-truth pose and symptom parameters remain available, enabling precise severity labels and auxiliary supervision.

Keywords: Computer Vision, Stable Diffusion, Healthcare, Generative AI, Video Analysis

Early motor development is one of the strongest indicators of a child’s neurological and physical health. Yet reliable assessment still depends largely on clinic visits or carefully staged recordings of conditions that rarely reflect how children actually move in everyday life. Homes introduce occlusions, arbitrary camera angles, motion blur, and partial visibility, while privacy concerns severely limit the collection of labeled data.

This project investigates whether modern generative AI can overcome these constraints by creating realistic, privacy-preserving training data. Students will design a simulation-driven pipeline that generates developmental motion and converts pose trajectories into photorealistic home videos with controlled difficulty. Models trained on this data will learn to estimate developmental age or milestone stage directly from RGB video, even when pose extraction is unreliable.

Keywords: Computer Vision, Stable Diffusion, Healthcare, Generative AI, Video Analysis

Military medics, police officers, firefighters, EMS crews, and pilots rely on highly structured voice protocols (e.g., MEDEVAC 9-Line, MIST, SBAR, ICAO distress calls, PAR reports) to transmit life-critical information under stress and in the presence of heavy ambient noise. These communications are increasingly captured as ASR transcripts, yet current NLP systems are not designed to reliably extract structured data from fragmented, error-prone radio speech.

This project develops and evaluates models that convert noisy ASR transcripts into structured incident fields (location, urgency, casualties, hazards, requests) across multiple standardized protocols. A synthetic data pipeline will be built to generate realistic training corpora: structured scenario generation → LLM-based spoken realization → radio noise injection → ASR simulation. Multiple extraction approaches (rule-based, fine-tuned LLMs, constraint-aware models) will be compared for robustness, hallucination control, and detection of missing critical fields.

Keywords: LLM, NLP, Voice, Mission Critical, Defense AI

Patient medication reviews often yield useful clinical insights, including symptom improvement, side effects, speed of action, and overall benefit–risk perception. However, datasets labeled at the symptom or attribute level are rare due to privacy limits and the cost of expert annotation.

This project builds an automated pipeline that converts formal drug documentation into a structured aspect ontology (e.g., symptom relief, nausea, sleep effects), then uses large language models to generate realistic, labeled synthetic reviews across different dosages, durations, and conditions. A multi-label transformer is trained to identify discussed aspects and assign sentiment to each one separately. The research tests whether synthetic data generated from controlled processes can support reliable, interpretable aspect-level sentiment analysis in healthcare, where real, labeled data is limited.

Keywords: LLM, NLP, Healthcare, Sentiment Analysis

Conventional medical AI systems are usually trained once on historical data using statistical or deep learning models. While they can achieve strong predictive performance, their knowledge is stored in opaque parameters and does not explicitly accumulate structured clinical reasoning. As a result, these systems do not gradually build interpretable expertise from experience the way clinicians do.

This project proposes an LLM-driven system that incrementally constructs clinical knowledge. By analyzing retrospective data, it identifies recurring thresholds, categorical interactions, and temporal patterns, and converts them into validated Python functions that represent explicit medical rules. Through a Chain-of-Code execution loop, each rule is tested, refined, and stored in a growing procedural knowledge base. The result is a complementary AI system that progressively builds transparent, executable decision logic from experience.

Keywords: LLM, NLP, Healthcare, Chain of Code, Vector Store, RAG