Building a generative face-replacement stage that swaps identity while preserving expression, gaze, and head pose across a 1M-recording, 750K-hour corpus, keeping the behavioral signal intact for training. Architecting the 8-stage anonymization pipeline across Whisper, pyannote, Presidio NER, and ARX k-anonymity, modeling compute down to a $25K to $165K core from a $0.9M+ baseline.

Built a retrieval-augmented generation system over 5,000+ clinical-supply documents, chunking and embedding them into a vector database with LlamaIndex for natural-language search across pharmaceutical PDFs. Added an ensemble OCR pipeline on Tesseract, PaddleOCR, and EasyOCR that pulls structured fields from messy labels and logistics forms, served through a Gradio chatbot for real-time document Q&A.

Contributing to a production open-source codebase through Social Summer of Code, one of the world's biggest open-source programs. 15+ pull requests merged so far, shipping features, bug fixes, and refactors reviewed by maintainers, plus issue triage and code review across the repo.

Built the Python annotation tooling and labeling workflow for a multimodal image-text misinformation benchmark, enforcing schema-validated labels and consistency across annotators. Engineered end-to-end dataset collection, validation, and versioning over 10,000+ image-text pairs into a reusable benchmark for evaluating vision-language models.

Shipped the React and Next.js front-end for the Trends platform, which reached 12,000+ unique users and 1,600+ active during peak course registration. Split heavy UI into route-level chunks and optimized asset delivery, cutting page load times by 35%.

Pushed Galaxy Zoo morphology accuracy +20% by fine-tuning VGG16/19 in TensorFlow with transfer learning and staged unfreezing. Built a repeatable ablation harness that ran 20+ controlled experiments across 6 unfreezing schedules and 10+ hyperparameter variants. Published in IEEE AIIoT with 15+ auto-generated plots.

Built an Instagram Q&A chatbot with Llama and Flask that answered event questions from posts, cutting manual lookups by 55%. Shipped the ingestion stack that ETL'd 2,000+ posts with BeautifulSoup, and mentored 100+ users on LangChain and Flask service patterns.

Ran AI/ML and cloud computing workshops for 40+ people. Helped 100+ club members get unstuck on Google Cloud, G Suite, and Android SDK, and mentored peers on LangChain, Flask microservices, and NLP integration.