SathwikMarupaka

A conversational AI that acts as a household's financial nervous system — grounded in real Plaid transaction and budget data, it reasons with Gemini to answer questions by checking live budget context, then orchestrates flight/hotel search, restaurant discovery, and books the loop closed with Calendar events, a Gmail PDF itinerary, and Twilio confirmations to the whole family. Built for the Personal Executive Agent hackathon track. (Co-built with Krishna Sai)

A pedagogical copilot that analyzes lecture video recordings to tell active teaching (whiteboard, gestures, demonstrations) apart from passive teaching (scrolling through slides). Samples keyframes and high-motion segments with OpenCV, uses Tesseract OCR to distinguish static slide text from skewed handwriting, and fuses frame + audio context through Gemini 2.5 Flash to produce a time-stamped engagement heatmap and an AI-generated executive summary for the professor.

IntelliLearn Studio uses a hierarchical multi-agent architecture — a SessionCoordinator delegates tasks to specialized sub-agents that analyze study materials, generate quizzes, and adapt difficulty (Easy/Medium/Hard) in real time based on student performance. It identifies knowledge gaps and surfaces targeted resources (e.g. relevant YouTube links) to close them, and persists student progress and history across sessions using Google ADK's state management.

Built at DevFest DC 2026's build-a-thon. Given any topic, generates three diagnostic questions engineered to expose misconceptions rather than test recall, then evaluates free-text answers to return a diagnosis — the single concept the learner's understanding actually breaks on — plus one concrete next step. Where a quiz returns a score, this returns a root cause. (Co-built with Nikhil Arethiya)

Forked and extended Google's Smart Buildings simulator to run a controlled SAC vs. PPO ablation — both agents trained in parallel on one GPU under identical compute budgets, optimizing a multi-objective reward balancing energy cost, carbon intensity, and ASHRAE comfort. Instrumented actor/critic loss straight from the TensorFlow graph to tie policy improvement to convergence rather than simulator variance. SAC outperformed PPO while completing 3x more training iterations.

Constructed an 855K-node / 1.96M-edge AST-derived knowledge graph (NetworkX) from a large Python codebase and benchmarked it head-to-head against a FAISS vector index over 455K Python files for LLM-assisted code generation. Evaluated both retrieval strategies on LiveCodeBench, raising Pass@1 from a 64.5% baseline to 84.25% with the stronger approach, and diagnosed why graph traversal underperformed in isolation (77.50%) — keyword-dependent lookup struggles to match natural-language problem statements.