02 · Generative AI · RAG

Ask-My-AI, a recruiter's assistant

A live assistant on artikumari.me that answers recruiters' questions about my work, grounded in my own documents and streamed as it thinks.

Role
Design and engineering
Platform
Web · artikumari.me
Year
2026
Stack
Next.js, Vercel AI SDK, Google Gemini, RAG, Embeddings, MCP
A laptop showing a calm chat interface answering a question
LiveAt artikumari.me
RAGAnswers grounded in documents
StreamedTokens arrive as they're written
01 · Context

Where this started

Recruiters skim a CV in seconds, then have the same questions: what did she build, with what, and is it real? A static page can't answer follow-ups. An assistant can, if it never makes things up.

02 · Challenge

What made it hard

  1. 01A general model will happily invent a job or a skill. That is worse than no answer.
  2. 02Answers had to feel instant on a personal site with no budget for heavy infrastructure.
  3. 03It had to be useful to someone who has never used a chatbot on a portfolio.
03 · Try it

Ask it something

A lighter version of the live assistant. It only knows the CV, and every answer names the section it came from.

Ask-My-AIDemo · answers from the CV only
Live version
Hi, I answer questions about Arti's work, using only her CV. What would you like to know?Source · Demo · grounded in CV
04 · Process

How it came together

  1. 01

    Collect the ground truth

    CV, project write-ups and the research summary become the only source the assistant may answer from.

  2. 02

    Embed and retrieve

    Documents are split and embedded. Each question pulls back the closest passages through vector search.

  3. 03

    Answer with Gemini, streamed

    Retrieved passages go to Google Gemini through the Vercel AI SDK, and the reply streams token by token into the page.

  4. 04

    Start the conversation

    Suggested questions sit under the input, so nobody faces an empty box.

05 · Decisions

Choices that shaped it

  • Grounded or silent

    If retrieval finds nothing relevant, the assistant says so and points to the contact page instead of guessing.

  • Streaming by default

    Showing the first words in well under a second matters more to perceived speed than total response time.

  • Agents and MCP where they earn it

    Tool calls are reserved for actions such as opening a project page, not for answering plain questions.

06 · Reflection
Building this taught me that the hard part of RAG isn't the model. It's deciding what the model is allowed not to know. Try the demo on this page, then the real one on artikumari.me.
Next case study · 03

Running construction sites from a phone

All work