My reflection
This project deepened my understanding of Human–AI Interaction from both sides of the interaction. While I wasn’t responsible for building the interviewer, developing an understanding of AI system architecture and flow, helped me work in the intersection, was also one of the most satisfying aspect from this research.
A personal learning for me came when things didn’t go as planned. Our recruitment pipeline stalled, and it was already close to finals week. I was tired, but I pushed and pivoted. And what surprised me I ended up enjoying it: the spontaneity, the little conversations, the unexpected curiosity from people who just met, connecting with people. (learning to enjoy the parts that don’t always follow the script)
This project also became the foundation for my first academic paper, which I co-authored as co–first author. We’ve submitted it for review, and while I’m hopeful it gets accepted, I’m already proud, not just of the research, but of the growth it represents ✨
Update: Paper accepted to ACM CSCW2025 🙏
Recommedations for Future work
Following usability testing, we identified key product opportunities that directly address usability gaps, emotional friction, and prep-stage flexibility. These recommendations focus on improving clarity, confidence, and control.
Dual Modes: Practice & Exam
Let users toggle between:
-
Practice Mode: guided, forgiving, repeatable
-
Exam Mode: timed, high-fidelity, no hints
Built-In Code Compiler
Users strongly expected to write and test code during the session. An integrated compiler would bridge the gap between explanation and execution.
Smarter Flow & Turn-Taking Cues
Introduce subtle visual or audio pacing signals to guide when to speak or listen, reducing hesitation and confusion.
Difficulty Level Settings
Participants wanted more control over challenge intensity. Adding adjustable difficulty (e.g., Beginner, Intermediate, Advanced) helps align the tool with users’ confidence levels and technical depth.
Visual Anchors for Focus
Incorporate timers, progress bars, or question counts to ground users and reduce uncertainty during sessions.
Mapping Current Landscape
Understanding the Need:
Coding platforms support technical practice (LeetCode, HackerRank), while human / peer-based mock interviews (Pramp, Interviewing.io) help simulate the interview experience.
But technical interviews require candidates to do more than reach the right answer. Students also need to:
-
Think and speak out loud while solving a problem
-
Articulate technical decisions in real time
-
Build confidence performing under interview pressure
-
Practice repeatedly before a real interview
yet realistic practice often depends on another person, which can introduce:
-
Scheduling and availability constraints
-
Unequal access to peers or mock interview partners
-
Added social pressure while still learning
The challenge wasn't simply getting more coding practice, it was getting opportunities to rehearse the cognitive, verbal, and interpersonal demands of the interview itself.
Turning the gap into Into Research Opportunity
Advances in generative AI created new possibilities for real-time dialogue, code evaluation, and multimodal interaction. But whether those capabilities could translate into a realistic and useful technical interview practice experience was still an open question. We focused our study around three research questions:
RQ1: Can an AI-driven mock interview realistically simulate a technical interview?
RQ2: How does interacting with an AI interviewer impact student confidence and perceived readiness?
RQ3: What are the benefits and limitations compared to traditional prep methods?
Our prototype: AI Mock Technical Interview System
To investigate our research questions, we built an AI-powered mock interview system combining generative models with multimodal interaction to simulate realistic technical interviews. It features:
-
A Unity-based frontend with a whiteboarding-style code editor
-
GPT-4o for dynamic interviewer dialogue and feedback
-
Deepgram (speech-to-text) and Silero (voice activity detection) for spoken input
-
Real-time code tracking and smooth verbal turn-taking (~300ms latency)
The interface mimics real interview conditions, minimalist by design, with only syntax highlighting and auto-indentation. Code and dialogue are continuously logged, enabling a structured yet low-pressure experience.
Key Takeaways & Impact: The 7 Core Themes Identified
Scroll right to left to continue viewing...
AI Felt Natural & Human-Like: This reduced intimidation and helped sessions feel more like actual interviews, without the social pressure.
01
Perceived Skill Improvement: Many students reported that the tool helped them practice speaking their thought process aloud, something they rarely get to rehearse, leading to increased self-awareness and pacing.
02
Realistic Practice, Not Just Simulation: The experience was perceived as close to a real technical interview, especially due to the conversational prompts and structured problem flow.
03
Confidence Boost Through Repetition: Several users shared that having a judgment-free, repeatable experience helped reduce anxiety and made them feel more prepared.
04
Issues with Pacing & Response Flow: Some participants struggled with the AI’s timing, feeling rushed, delayed, or unsure when to respond, which disrupted focus.
05
Desire for Visual Support & Feedback: Users wanted progress indicators, time cues, or light guidance to manage expectations and stay grounded during the session.
06
Need for Personalization: Students wanted more control over difficulty, pacing, and AI tone, preferring a tailored experience that aligned with their interview stage
07
Service Name
This is the space to introduce the Services section. Briefly describe the types of services offered and highlight any special benefits or features.
Service Name
This is the space to introduce the Services section. Briefly describe the types of services offered and highlight any special benefits or features.
Service Name
This is the space to introduce the Services section. Briefly describe the types of services offered and highlight any special benefits or features.
🎯 Evaluating the Human-AI Interview Experience
With our participant pool locked in, I conducted moderated testing to understand how students experienced the alpha prototype, not just in terms of system-level usability, but through participants’ mental models, affective experience, and behavioral responses and the conversational behavior of the AI itself.
Survey
-
Before testing, participants completed survey capturing:
-
Major + academic level
-
Past technical interview experiences
-
goals & pain points
-
Demographic context
-
🎯 Target Audience for our Study
In defining who the tool should serve in its earliest iteration, I focused on identifying a specific, high-need user group that would benefit most from solo, verbal technical interview practice.
While the broader market for interview prep includes everyone from bootcamp graduates to senior engineers, our Alpha prototype needed to start with a more focused segment, one with clear friction and unmet needs.
📍 Immediate Reach:
To test early assumptions, I focused on upperclassmen in computer science at Drexel University, specifically juniors, seniors, and graduate students preparing for technical internship or entry-level roles. This group became our initial prototype testing audience.
Recruiting Candidates...
While working under Dr. Tiffany Do, our original plan was to recruit through Drexel’s internal system, mass emails sent by CCI admin staff. But like many institutional processes, it moved at snail space 🐌. With just 2–3 weeks reserved for user testing, we couldn’t afford to wait.
So I pivoted 🧭 Plan B:
I went grassroots! walking through CS buildings, lounges, and labs with my elevator pitch ready. It almost felt like being a real estate agent cold door-knocking. No scripts, just energy, adaptability, and purpose!
It pushed me out of my comfort zone, I had to be proactive, clear, and quick on my feet. I even made a few new friends in the process 🤭
Within those two weeks, I personally recruited 11 students who matched our core user segment, and their feedback shaped everything that followed.
🎙️ Live Session + Post-Interview
articipants completed a real-time, whiteboarding-style technical interview with the AI prototype, working through the same unseen coding problem while verbalizing their thought process.
Each session was followed by a semi-structured interview exploring their experience.
Analysis & Thematic Synthesis
During each session, I paid attention to both sides of the Human ↔ AI interaction—how participants responded to the interviewer and how the AI’s behavior shaped the exchange as it unfolded. I focused especially on moments where the two intersected.
Human side
-
Behavioral signals – pausing, repetition, filler words
-
Mental models – how they interpreted AI prompts, responded to questions, or moved through the task
-
Affective Experience – visible nerves, moments of ease, confidence dips or post-session relief
AI system design
-
Turn detection & speech boundaries — interruptions, premature responses, or extended waits associated with how the voice agent detected conversational turns
-
Response latency — participant-perceived delays between speaking and receiving an AI response
-
Contextual responsiveness — whether AI responses appeared to follow participants’ verbal reasoning and evolving code
-
These interaction moments, alongside participants’ post-interview reflections, became inputs into our analysis.
Researching Confidence in Technical Interview Prep
Undergrad HAI Research Assitant @ Diva Lab, Drexel University SCIS | Study duration: Dec 2024 – Apr 2025
This project explores how AI can support students in practicing technical interviews in a more human, conversational way. Most platforms focus on solving problems, but not on articulating thinking under pressure. We investigated a designed tool that simulates real technical interviews through an AI voice agent, giving students low-pressure, repeatable practice that feels real.
The 7 Core Themes Identified
Service Name
This is the space to introduce the Services section. Briefly describe the types of services offered and highlight any special benefits or features.
Service Name
This is the space to introduce the Services section. Briefly describe the types of services offered and highlight any special benefits or features.
Service Name
This is the space to introduce the Services section. Briefly describe the types of services offered and highlight any special benefits or features.
🗂 Using Braun & Clarke’s six-phase framework
We analyzed session transcripts and post-interview responses using Braun & Clarke’s six-phase thematic analysis framework, we moved from transcript familiarization and initial coding to grouping and refining patterns across participant responses. Through collaborative coding and discussion, we identified 42 codes which were iteratively synthesized into 7 higher-level themes.
75% Said the AI Felt Realistic
Participants were surprised by how natural the AI interviewer sounded especially in tone, pacing, and prompt delivery
78%Confidence Boast
The judgment-free format helped reduce anxiety and made students feel more prepared to speak in real interviews.
80% Found the Tool Useful
The vast majority of participants described the experience as helpful, realistic, and aligned with how they want to prepare
IMPACT
Framing the Problem
Technical interviews test more than whether someone can solve a problem, they also require candidates to articulate their thinking and navigate a live interaction.
