AI Twin
A glass-box RAG chatbot, grounded in a curated corpus.
My AI twin answers questions about me and my work, grounded only in a curated corpus. It is the same hybrid-RAG pattern I build professionally, pointed at myself, and it shows you the exact chunks behind every answer.
How this works
The twin is deliberately over-engineered for a corpus this small, because the pipeline is the point.
- Chunking. The corpus is plain markdown, one topic per heading. A deterministic chunker splits it into stable, retrievable chunks at build time.
- Hybrid retrieval. Every question runs through two retrievers at once: BM25 over the keywords, and cosine similarity over Gemini embeddings. Reciprocal Rank Fusion merges the two rankings, so exact-term questions and paraphrased ones both land.
- Grounded generation. The top chunks go into the prompt as data, not instructions, and Gemini answers in my voice under strict rules: corpus only, no invented facts, decline what it does not know.
- The trace. The panel under each answer shows the real chunks, their rank in each retriever, and the fused score. No black box.
What it will and won't do
It knows my projects, the paper, my experience, and roughly what I am looking for. If a question is outside the corpus, it says so and points you to email rather than guessing, and it declines off-topic or adversarial prompts in character.