Nishad Patne

Projects / Digi Twin

Completed

Digi Twin: a digital twin for my ex‑portfolio, a hybrid‑retrieval chatbot

1 / 7The assistant on my previous portfolio.

Product
np-rag, the portfolio AI assistant
Role
Design and engineering
Team
Nishad P
Timeline
March 2025 to March 2026
Status
  • Retired
  • switched off with this portfolio's launch
Built with
  • Python
  • Haystack
  • Qdrant
  • FastAPI
  • SPLADE and OpenAI embeddings
  • Gemini
  • Oracle Cloud
  • Vercel

A chatbot that answered visitors' questions about my work, like a digital twin.

Shown at LEI and in interviews; it worked very well for at least a year.

01

Concept and work

  • WhyI wanted to learn AI, and no free, out-of-the-box option existed then.

  • Two modesGeneral, about me, and Project Deepdive, on each case study.

The assistant on my previous portfolio.
02

The interface

  • Built around the assistantI vibe-coded the portfolio so the chatbot felt like part of an app: a side panel, starter questions, and an "Ask AI" button on each case-study section.

  • Why vibe‑codedWebsite builders like Framer, Webflow, Wix or Squarespace wouldn't take custom code then; agentic platforms like v0 would. Since AI, most builders allow code injection too, and their value has shifted toward templates.

In use.

1 / 3Project Deepdive, summarising a case study
03

How it answers

  1. 01

    A gatekeeper screens out off-topic or gibberish questions.

  2. 02

    A retriever searches my portfolio by exact words (SPLADE) and by meaning (OpenAI embeddings), and ranks the results.

  3. 03

    A writer, Gemini, answers from those passages, so replies stay grounded in my work.

  4. 04

    A session manager, built in Python with Haystack, keeps each visitor's chat separate, so follow-ups make sense.

No dead ends

Off-topic questions get a redirect, gibberish gets a request to rephrase, and a miss comes with my email (and a GIF).

The RAG pattern it's built on: retrieve context, then answer from it.
04

Learning and pivots

  1. Started fully local (Ollama, LangChain, Chroma) to learn every piece hands-on, until my laptop ran out of compute; generation moved to Google's API, storage to Qdrant.
  2. Went hybrid: kept SPLADE for exact words, added OpenAI embeddings for meaning.
  3. Moved to Oracle's free tier, because I couldn't keep my laptop on 24×7.
  4. Retired when updates got cumbersome, my data went stale and the free tiers kept shrinking.
1 / 3My learning trail in ChatGPT, one step at a time.
05

Testing

Manual conversations and automated simulations, including several visitors at once, exposed sessions bleeding into each other under load. A single server worker and new session clearing fixed it.

06

Reflection

Looking back, this was my early exploration of agentic UX, and I can see how much digital expectations have changed, and how far LLMs have come since I built it. It taught me how agents work, what orchestration, vector databases and embeddings are, and how an orchestrator keeps context. I co-wrote it with Codex, ChatGPT and Claude, and I can explain every piece of the code.

Worthy of a CS grad student's project; for me, massive learning, and I used it very dearly.