AI-POWERED INTERVIEW ASSISTANT

Dr. R Seeta Sireesha

Department of ComputerSciene and Engineering
Gayatri Vidya Parishad College of Engineering (Autonomous)
Visakhapatnam
[email protected]

Rameez Ahmad Jalagadugula Bhavya Sri Pakanati Vijaysimha Reddy Nuzhath Tahseen
Departmet of CSE (AI & ML) Departmet of CSE (AI & ML) Departmet of CSE (AI & ML) Departmet of CSE (AI & ML)
GVPCOE (A) GVPCOE (A) GVPCOE (A) GVPCOE (A)
Visakhapatnam Visakhapatnam Visakhapatnam Visakhapatnam

Abstract— In today's dynamic educational environment, technical students juggle coursework, internships, and jobs, making
skill assessments challenging. As graduation nears, inadequate interview preparation can cause anxiety and erode confidence.
Rigid schedules and fixed interview dates hinder students' ability to link theoretical knowledge with practical skills needed in
the competitive tech industry. Current models lack personalized interview experiences, leaving students without proper
guidance. Our solution leverages OpenAI keys and LangChain models to revolutionize interview preparation by generating
tailored questions from resumes. Streamlit enables smooth interaction, while OpenAI integration enhances simulation
sophistication, bridging theory and practice. This comprehensive approach empowers students to excel confidently in job
interviews.

Keywords— Interview, LangChain , OpenAI , Streamlit.

I. INTRODUCTION
In the realm of career advancement, meticulous interview preparation emerges as a cornerstone of success.
This innovative application, powered by cutting-edge technologies such as Streamlit, LangChain, and
OpenAI, epitomizes a sophisticated solution tailored to meet the evolving needs of today's job seekers. By
seamlessly integrating advanced language processing mechanisms, it offers a refined approach to
interview readiness, empowering individuals with personalized insights and guidance. At its essence, the
application embodies efficiency and efficacy, leveraging the prowess of Streamlit for intuitive user
interaction and LangChain for seamless text processing[5][6]. Through the lens of OpenAI's language
models, it navigates the complexities of resume parsing, swiftly distilling pertinent information and
crafting tailored interview questions[2]. This symbiotic integration of technology not only optimizes the
preparation journey but also ensures that candidates are equipped with a comprehensive understanding of
the topics they may encounter during interviews. Furthermore, the project objective transcends mere
question generation; it aspires to foster a culture of continuous improvement and empowerment. By
facilitating audio recording capabilities and leveraging LangChain's capabilities, it enables candidates to
articulate their responses with clarity and precision[5]. Through iterative analysis and feedback loops, the
application empowers individuals to refine their communication skills, ultimately enhancing their
confidence and competitiveness in the job market.

II. METHODOLOGY
The primary objective of this application is to revolutionize the interview preparation process by offering a
comprehensive solution that combines technological innovation with strategic foresight. By integrating
Streamlit[6], LangChain[5], and OpenAI[2], the application aims to provide a tailored and immersive
experience that surpasses traditional preparation methods. To streamline the preparation journey, the
application leverages Streamlit's interactive interface, ensuring user-friendly navigation and
engagement[6]. This interface facilitates a seamless preparation journey for users, making the process
intuitive and efficient. The focus on interactivity and ease of use is designed to reduce the complexity
often associated with interview preparation tools, thereby enhancing user satisfaction and effectiveness.
The application utilizes LangChain's advanced text processing capabilities to automate the extraction of
relevant information from resumes[5]. This involves parsing the resume to identify key skills, experiences,
and qualifications, and then categorizing and organizing the extracted information for further processing.
Based on the analyzed resume data, the application generates personalized interview questions that align
with the candidate's unique profile. This personalized approach ensures that the preparation is relevant and
targeted, addressing the specific needs and strengths of each individual. To enhance candidates'
communication proficiency, the application incorporates audio recording functionality. This enables users
to articulate their responses verbally, which are then recorded and stored for further analysis. By
leveraging Google's speech-to-text conversion capabilities[3] and text-to-speech conversion capabilities[4]
the application transcribes these verbal responses into text, allowing for detailed analysis of the content
and delivery. This feature provides a dynamic and immersive preparation experience, fostering better
communication skills through practical application. The application performs an iterative analysis of the
recorded responses, evaluating the accuracy, clarity, and relevance of the responses. Based on this
analysis, the application provides constructive feedback to users, helping them refine their responses and
improve their overall interview performance.

III. EXISTING SYSTEM

OpenAI's high API usage costs and ethical concerns regarding biases in question generation may hinder its
suitability for large-scale interview preparation[2]. Similarly, platforms like Gemini may lack the
necessary customization, while Hugging Face's models might require complex integration and lack
specialized capabilities, diverging from the project's objectives. Brad.ai's focus on coaching may not align
with automated question generation goals, and 1:1 mock interview could lack scalability compared to
automated systems. Furthermore, concerns about the cost and ethics of OpenAI's language models persist,
while Gemini's primary focus on scheduling may restrict customization options[2]. Integrating Hugging
Face's models poses complexity and may lack the specialized capabilities needed, and Brad.ai's coaching
emphasis does not align with the project's aims. Moreover, 1:1 mock interview lacks scalability compared
to automated systems, making them less suitable for widespread implementation.
Drawbacks of the Existing System: The Traditional hiring struggles with PDFs and manual review.
Parsing complex PDF layouts hinders data extraction for applicant tracking systems (ATS). Manually
reviewing large applicant pools is inefficient, and human bias can lead to inconsistent decisions.

IV. PROPOSED SYSTEM

In our algorithm, we aim to develop a robust system for efficient document retrieval and processing,
leveraging advanced techniques such as document loaders, text splitting, embedding models, vector stores,
retrievers, and indexing.
LangChain : LangChain as shown in Fig-1 is a robust framework designed for developing applications
powered by language models. It provides essential tools and components for efficiently managing and
processing text data[5]. Key features include text parsing, natural language understanding, and advanced
text generation capabilities. LangChain supports seamless integration with various APIs and data sources,
making it highly versatile for building custom language-based applications. It excels in automating
complex text analysis tasks, such as extracting relevant information from large datasets and generating

coherent, contextually appropriate responses[5]. This makes LangChain a valuable asset for developers
seeking to create sophisticated, language-driven applications.

Document loaders : Document loaders act as the primary entry point for bringing data into our system.
They provide the initial step in the data ingestion process, facilitating the seamless integration of textual
content from various sources[1].
Text Loader : The Text Loader component serves as a foundational element in our system, responsible
for sourcing textual documents from various data repositories[1]. By seamlessly interfacing with diverse
sources including local files and cloud-based storage solutions, Text Loader ensures the reliable
Fig-1 Langchain Architecture
acquisition of data essential for subsequent processing and analysis.
Unstructured URL Loader : The Unstructured URL Loader expands our system's capabilities by
enabling the retrieval of unstructured data from web sources[1]. Through sophisticated web scraping
techniques, this component facilitates the extraction of information from publicly accessible URLs,
enriching our dataset with external content for comprehensive analysis and insight generation[1].
Character Text Splitter : At the core of our data preprocessing pipeline, the Character Text Splitter
module plays a pivotal role in segmenting large textual documents into manageable fragments[5].
Utilizing sophisticated character-based splitting algorithms, this component optimizes data processing
efficiency and enhances retrieval performance by isolating relevant sections of text.
Recursive Character Text Splitter : Building upon the capabilities of its predecessor, the Recursive
Character Text Splitter further refines the text segmentation process through recursive parsing techniques.
This advanced algorithm ensures precise extraction of meaningful content from complex documents,
facilitating accurate representation across diverse formats and structures.
Streamlit UI : The Streamlit UI component serves as the user-facing interface of our system, providing
intuitive access to its functionalities[6]. Designed for simplicity and ease of use, Streamlit UI enables users
to explore, query, and visualize data effortlessly. By offering a seamless and interactive experience, the UI
enhances user engagement and ensures efficient utilization of our system's capabilities across diverse
applications and use cases[6]. Built upon Streamlit's framework, the UI offers a user-friendly experience,
enabling effortless access to various functionalities and insights. Concurrently, project coding
encompasses the implementation of underlying algorithms and logic, ensuring the robustness and
functionality of our system[6]. Through meticulous coding practices and adherence to best practices, we
uphold the integrity and reliability of our solution.
Text Splitters : Text Splitter as shown in Fig-2 efficiently breaks down large documents into manageable
chunks, enhancing processing efficiency and enabling targeted analysis[5]. Coherent Chunking utilizes
advanced algorithms to ensure that text chunks maintain coherence and relevance, preserving the
contextual integrity of the original document. Optimized Processing by segmenting text into smaller units,
Text Splitter optimizes subsequent retrieval and analysis processes, facilitating faster and more accurate
information extraction[1].

Fig-3 Vector Database

Fig- 2 Text Splitters

Vector Database : In the ever-evolving landscape of artificial intelligence, vector databases as shown in
Fig-3 stand as pivotal solutions, indexing and storing vector embeddings to enable swift retrieval and
similarity searches. As we navigate through the AI revolution, these databases emerge as indispensable
tools, addressing the escalating complexity and scale of modern data processing. By harnessing the
semantic richness embedded within vector representations, they empower applications reliant on large
language models and generative AI, facilitating efficient knowledge retrieval and long-term memory
maintenance. Through seamless with embedding models, these databases augment AI capabilities,
facilitating tasks such as semantic information retrieval with unparalleled efficiency[5]. Thus, they play a
pivotal role in enhancing the effectiveness of AI-driven applications, embodying the synergy between
advanced data management and transformative AI innovation.
FIASS : FAISS (Facebook AI Similarity Search) as shown in Fig-4 is a cutting-edge library designed for
efficient similarity search and clustering of high-dimensional vector data. Developed by Facebook AI
Research, FAISS offers optimized algorithms tailored for large-scale datasets encountered in AI
applications[7]. Its advanced indexing techniques, such as Product Quantization (PQ) and Hierarchical
Navigable Small World (HNSW), ensure rapid and accurate nearest neighbor search operations. FAISS
supports essential functionalities like CRUD operations and metadata filtering, simplifying data
management. Additionally, FAISS enables horizontal scaling, distributing index structures across multiple
machines for enhanced performance and scalability.

Fig-4 FAISS Indexing

Retrieval : Retrieval mechanisms as shown in Fig-5 orchestrate the process of fetching relevant data
based on user queries, bridging the gap between raw data and actionable insights[7]. The
RetrievalQAWithSourcesChain leverages sophisticated algorithms to identify and retrieve pertinent
information, taking into account multiple data sources and query types[7]. By employing techniques such
as semantic search and ensemble retrieval, it enhances the precision and comprehensiveness of search
results, empowering users with actionable knowledge.

Fig-5 Retrieval

Retrieval Questions and Answers With Sources Chain : The RetrievalQAWithSourcesChain module
represents the pinnacle of our system's retrieval capabilities. Incorporating advanced algorithms, this
component enables users to pose complex queries and retrieve relevant documents with exceptional
efficiency[5]. By integrating multiple data sources and leveraging semantic understanding,
RetrievalQAWithSourcesChain empowers users to extract actionable insights from vast repositories of
textual data with unparalleled accuracy and speed.

V. RESULTS AND DISCUSSION

The application operates by first receiving user input. It then processes this input using algorithms based
on the proposed system's design. Next, it executes the necessary actions, generates output, and presents it
to the user. Finally, it may prompt for further interaction or loop back to receive additional input. The
entire workflow of the application is shown in Fig-6.

Fig 6 Workflow

This is how the proposed system helps the user to upload their resume and generate interview questions,
providing them with an interface for recording their audio and then analysing it for improvement in their
answers, ultimately leading to better performance in actual interview sessions.

REFERENCES

[1] Data Extraction from pdf: https://pypi.org/project/PyPDF2/

[2] Open Ai docs: https://platform.openai.com/docs/introduction
[3] Google Cloud Speech to text API: https://cloud.google.com/text-to-speech?hl=en
[4] Google Cloud text to Speech API: https://cloud.google.com/speech-to-text?hl=en
[5] LangChain docs: https://python.langchain.com/docs/get_started/introduction
[6] Streamlit : https://docs.streamlit.io/
[7] FAISS : https://faiss.ai/index.html