An AI stand-in that answers recruiters’ questions about me over a phone call or chat, grounded in my résumé and GitHub, and books real interviews on my calendar with no human in the loop.
Play demoThe chat interface. The same agent also answers a real phone line.
Python
FastAPI
LangChain
Vapi
ChromaDB
What it does
Built for the Scaler AI Engineer screening assignment. A recruiter can phone a number or open a chat page and ask about my projects, skills and availability. The agent answers from what I have actually done, then checks my Google Calendar and books the interview itself.
How it works
Voice runs through Vapi: Deepgram transcribes the caller and Azure voices the reply, while my FastAPI server stands in as the language model behind an OpenAI-style /chat/completions endpoint.
A LangChain agent on GPT-4o with tool calling decides whether to answer, check my availability or book a meeting through the Google Calendar API.
Answers are grounded in 132 chunks from my résumé, GitHub READMEs, repo summaries and commit history, stored in ChromaDB and retrieved fresh on every turn, so a follow-up question still finds the right context.
Spoken email addresses (“name at the rate of gmail dot com”) are cleaned up before they reach the calendar invite.
How well it works
Measured on 28 golden questions scored by a GPT-4o judge, plus live calls:
About 1.4 s from the end of a question to the start of the answer (about 400 ms of that on the server).
Bookings succeeded 4 times out of 5.
Hallucinated on 2 of 28 questions (7.1%) and resisted 4 of 4 prompt-injection attempts.
Average answer quality of 3.96 out of 5.
A five-minute call costs about $0.11 to run, and a ten-message chat about $0.05.
Play demo