Intervuew(Real_time interview Agent) with RAG

Intervuew is a real time interview agent for tech rolees. that can conduct interviews, has real time audio interaction. it helps organizaitons reach more people and automate hiring process

Comment

What it does

Intervuew is an AI platform designed to aid the hiring of tech recruits, and can also be used as a learning and preparation tool for applicants. It communications with user using a live session, gives a live transcript of the user and AI audio, after the interview. it scores the interviewee based on technicality, accuracy and exceptionality. Then proctoring detection system with video analysis, will detect if the interviewee cheated or broke the rules while taking the interview The key features are:

1. Real-time Audio: The user communicates with the AI via audio and a live transcript is sent and shown during the interview. The AI acts as the interviewer agent that askes question based on retrieved question data from the Vector Search(RAG), and also might also ask the user to elaborate or give more details on his/her answer. The interview round is 10 minutes, and warning is sent 10 minutes before it closes. The Interview agent has it's own persona. once the session ends, the evaluator agent scores the applicant

2. Proctoring and Cheatinng: After the live session has ended, all frames from the interview are collected and gathered for review, and then sent to an agent that will check if there is any signs of cheating and send a report and analysis of those frames back.

3. AI-Generated Questions with Retrieval Augmented Generation(RAG): the questions are gathered into a vector databases after being collected from relevant and free sources. The AI uses the vecor database in an efficient function call to get the relevant question data based on what is set in the interview. Questions are also generated from the applicants resume.

How we built it

We built the model by leveraging both google adk and sdk for real-timie audio communication and transcription, and sdk for post video processing

Real time audio model: gemini-2.5-flash-native-audio-latest for real-time audio communcation and live transcription. implemented caching and context window compression to reduce load of the initial request

Proctoring video analysis:R eportgemini-2.5-flash for post proctoring report and evaluating interview. for porctoring report it uses Frame ingestion with concurrent frame loading(i..e the images in the frame are concurrently loaded to a video file). and workers that work as background process. so the proctorin event doesn't block other processes.

API: FastAPI was implemented as the api for handling the request.

Database: Sqlite3 deployed on google-cloud sql using postgresSQl..

RAG implementation: Qdrant for the vector database, and broke down each query so the search would be efficient.

Front-end: Typescript and React

Error Handling: Processes is rebatched and sent back for analysis and evaluation, even after experiencing an error. so it will be evaluated and recorded

Challenges we ran into

The major challenge of this project was the limit of using a low tier access, i was therefore limited by ratelimits and had to optimize my model around the ratelimits using: context-caching and window-compression.

What's next for Intervuew(Real_time interview Agent) with RAG

Multiple applicants we plan on creating a the same system for multiple meetings, where diffrent, AI interviewees can interact and ask questions, and also one or more applicants

Adding more sentiment Next focus is on making the AI sound more human and approachable.

Built With

Share this project:

Updates