Projects
What I built, and how I evaluated it.
Selected work across applied AI, software engineering, computer vision, and multimodal learning.
Clinical document QAOngoing personal project
CareMind In progress
Problem: Answering questions across medical documents requires choosing the right workflow and retrieving supporting evidence.
PythonFastAPILangGraphSupabaseRedisNVIDIA NIMMCPNext.jsTypeScript
Explore project +
My work: Built LangGraph routing for retrieval, report comparison, and medical education. Implemented MCP tools for search, comparison, and timeline extraction; connected NVIDIA NIM embeddings and Supabase vector search to cited answers.
Integrated a Next.js frontend, FastAPI backend, and Redis response cache. Built a repeatable evaluation pipeline for route accuracy and citation pass rate.
Scope: Uses synthetic medical PDFs. Evaluation is ongoing; no clinical deployment or patient outcome claim.
Document processingFull-stack application
Cognito-Map
Problem: Questions embedded in PDFs are difficult to extract and organize for educational analysis.
FlaskReact.jsMongoDBPineconeChart.jsTypeScriptGemini API
Explore project +
My work: Built an ETL pipeline with 94% PDF question extraction accuracy using Gemini, stores them in MongoDB, and makes them available through a Flask and React dashboard with Chart.js visualizations.
Result: Compared five models for Bloom’s Taxonomy question classification on the same test set. The Sentence Transformer achieved 78.94% accuracy, outperforming the strongest baseline, KNN, by 10.85 percentage points.
Computer visionMedical imaging
Prostate MRI super-resolution
My work: Developed a multi-image pipeline for reconstructing missing or low-resolution prostate MRI slices. Compared CNN, SRGAN, and diffusion approaches.
PyTorchCNNSRGANDiffusion ModelsOpenCVNumPy
Explore project +
Evaluation: Used PSNR, SSIM, and qualitative anatomical assessment to examine the tradeoff between image sharpness and structural fidelity.
Multimodal learningAudio + text
Spoken command recognition
My work: Combined Wav2Vec2 audio embeddings and BERT text embeddings using cross-attention and Transformer decoders. Added Pinecone retrieval to select semantically relevant text commands.
PyTorchHugging FaceWav2Vec2BERTPineconeScikit-learn
Explore project +
Evaluation focus: Compared fusion strategies and investigated generalization under noisy audio conditions.
Learning assistantHackathon project
WaitWhat?
My work: Built a video learning assistant that uses facial-expression signals to flag possible confusion and record timestamps. Integrated Whisper transcription, BART summarization, and OpenAI-powered question answering in Gradio.
PythonGradioOpenCVPyTorchTensorFlowTransformersWhisperBARTOpenAI APIyt-dlpMoviePyYouTube API
Explore project +
Scope: Supports YouTube and local video. Facial expressions are a proxy for possible confusion, not a definitive measure of understanding.