Demo

Interactive demos and experiments. The current feature is a RAG assistant that answers questions about my work and returns sources.

RAG demo CPU mode

Ask about my work

Portfolio Q&A demo with retrieval + citations.

Hi — ask me about my research, projects, or publications. I’ll answer using my site notes and show sources.

Enter to send • Shift+Enter for a new line

Demo note

This demo currently runs on CPU, so responses can be slower than a GPU-backed setup. I’m prioritizing GPU resources for experiments/training and research workflows. When spare GPU capacity is available, I’ll enable faster GPU demos again.

Tip: short questions feel much faster on CPU.

How this demo works

Open

Retrieval-Augmented Generation (RAG): we retrieve relevant notes, then the model writes an answer grounded in those sources.

RAG flow diagram: user question -> retriever -> LLM -> answer with sources.

History

Recent turns are included so follow-ups stay consistent.

Retrieval

The backend searches notes/docs and injects top snippets as context.

Generation

The model generates tokens step-by-step to form the final answer.

Roadmap: agent mode

This demo will evolve from “answer questions” to “run my workflow.” Instead of one generic chatbot, an agent will route each request to the right capability: retrieve context from my notes (RAG), run my vision models (e.g., segmentation/detection/embedding search), or run analysis + generate visuals (plots, metrics, evaluation slices).

Then it will combine outputs into one clear, grounded explanation—so visitors can see how I actually work across remote sensing, medical imaging, and data science through a single interface.

Open

Agent flow diagram: planner routes to tools (RAG, image model, data viz) then produces a grounded response.

Why agent mode

Tool routing
Decide whether to retrieve notes, run a vision model, or generate a plot.
Bounded steps
Limits + timeouts keep it stable (especially on CPU).
Grounded answers
Cite sources or say “not sure” when info isn’t available.