Div Agarwal

Problem: Neuroscience teams need to inspect calcium-trace events quickly without forcing every experiment through a slow desktop workflow.

Approach: I built a browser-based analysis interface in Next.js and TypeScript that makes neural traces explorable, shareable, and easier to pair with downstream event-detection research.

Result: The surrounding research workflow reached a 0.92 F1 event-detection score, and the tool ships as a live demo for interactive inspection instead of static plots.

Problem: Long expert-call transcript archives are difficult to search, summarize, and cite reliably when every answer has to stay grounded in source material.

Approach: I built a local RAG stack with Llama 3.1, LlamaIndex, LanceDB, and a Gradio interface so retrieval, inference, and source citation all stay modular and replaceable.

Result: The system runs fully locally, returns source-grounded answers over transcript collections, and the current README documents roughly 1-2 minute response times on local hardware.

Problem: Content publishing breaks down when image selection, captioning, search, and posting still require manual handoffs across tools.

Approach: I designed a serverless AWS pipeline that connects image retrieval, Rekognition analysis, Claude-based captioning, OpenSearch indexing, and publishing logic into one automated flow.

Result: The project automates a 5-stage content pipeline end to end, reducing the workflow to one orchestrated system instead of separate manual steps.

Problem: Collections teams need a usable operating surface for turning messy account data into actionable outreach plans rather than static spreadsheets.

Approach: I built a full-stack product with FastAPI, React, SQLite, and the OpenAI API to generate strategies, manage contacts, and organize actions in a timeline-based UI.

Result: The system handles both Excel and PDF imports, then turns that data into AI-generated multi-step strategy blocks inside a single workflow interface.

Problem: A useful shared knowledge assistant has to support multiple active users without cross-session context bleed or fragile state management.

Approach: I designed a session-aware RAG architecture that separates shared retrieval infrastructure from per-user conversational state so multiple users can query the same corpus safely.

Result: The system supports concurrent multi-user access over a shared knowledge base while keeping retrieval context isolated at the session level.