Landing Page for AutoRecruit
Each Interview pertains to a particular job, and can be associated to multiple sessions with different transcripts and scores!
Interview Dashboards
Sample interview UI
Question Transcript during the live interview
Generated summary from Job description and linked interview sessions associated with an interview board
Sample transcript from the Interview. 'Initiate analysis' outputs a detailed summary based on custom context-mapped real-time conversation.

AutoRecruit

Prepare for your next big job interview like never before with AutoRecruit, the AI-powered Behavioral Interview Platform. AutoRecruit offers real-time, tailored mock interviews that replicate real-life interview scenarios. Through our comprehensive post-interview analytics, identify your strengths and improve on your weaknesses. Equip yourself for success and conquer the job market with unmatched preparation!

Inspiration

Navigating the job market is a daunting task, and interviews are often the most nerve-wracking part of the journey. AutoRecruit is designed to empower job seekers by providing them with a lifelike interview experience. This way, users can walk into real interviews armed with valuable insights and confidence, transforming the interview from a challenge into an opportunity.

What it does

Django: Back-end framework for robust application handling
React: Front-end development to deliver an interactive user experience
Redis: In-memory data structure store for caching and real-time analytics
Deepgram: Speech recognition API for real-time transcription during interviews
Modal: Rust-based fully-managed container system for quickly scaling our AI models
TTS (Text-To-Speech): To provide real-time spoken prompts and feedback during interviews
ElevenLabs: ASR system for converting speech into text
OpenAI: Integrated for advanced natural language processing capabilities

How we built it

The individual components that enable AutoRecruit are as follows:

Dynamic prompt-engineering: We used GPT-4's enhanced conversational capabilities to generate questions on the fly, keeping in mind the context of the interview in real-time and the keypoints in the job description that is missing from the applicant's resume. We also consider situations were the recruiter might want to inform the applicant about a potential mismatch, and thus weigh our skills and qualifications highly. We have a context-map of questions that need to be answered for the recruiter (now, the system) to evaluate the applicant completely. This enabled a smooth flow between the interview and interviewee, making it seem more of a human-to-human conversation.
Custom evaluation metrics: The interview and evaluation process in real life is a very subjective one, and we decided to go the forego the route of generating a fixed-set of objective metrics. We targetted a unique metric system that uses the emotions of the applicant while answering questions and the confidence in their voice to enhance how strong their answers are, apart from just being able to speak the most important keywords. We fine-tuned a Multi-layer perceptron for the audio-analysis and DeepFace for emotion-detection. We deployed custom endpoints using Modal.
Real-time interviews: We integrated our React front-end with Django back-end with WebSockets for consistently transcribing audio without the need to chunk and store the audio. This was the most important aspect of AutoRecruit, and was the hardest to implement.
Job-description-based Resume building: Based on every Interview Dashboard, the applicant's resume is parsed and suggestions are generated for improving the resume specifically for the job. This allows the applicant to fine-tune their resume for every job they want to apply for!

Challenges we ran into

The most daunting challenge was the seamless integration of all these diverse technologies into a coherent, functioning platform. Specifically, we struggled to get our custom LLM interview pipeline to work in tandem with the other components, which was crucial for the assessment aspect of the application. Real-time data synchronization across different tech stacks and ensuring a secure and efficient data flow were other roadblocks we encountered.

Accomplishments we're proud of

Even though we were not able to integrate all the features that we worked on, we are very happy with what we were able to accomplish in a span of 36 hours. If we had more time, we believe this project could be transformed into a real-product capable of helping the entire college community.