Inspiration
Standardized test prep is high-stakes and high-cost, yet still relies on passive methods that struggle to hold attention, create accountability, or guide students in real-time. In a world where students constantly juggle notifications, apps, and distractions, dense textbooks and pre-recorded videos often fail to keep learners engaged when it matters most. Moreover, 1:1 tutoring/counseling is expensive for these exams. From SAT/ACT, GMAT, and MCAT admissions to professional exams like the CFA, BAR, and SIE, families and individuals spend thousands on prep that doesn’t adapt in the moment.
Having experienced these gaps firsthand and knowing I’ll face them again throughout my career, I wanted to learn how to study and keep myself accountable.
What it does
Most AI tutors are just ChatGPT clones. Mine uses psychology-backed study methods (Pomodoro (short, intense study blocks) and Feynman Technique (active recall)) and voice-first design to feel like a real "LIVE" tutoring session working through standardized test materials. Students don't just get answers, they learn HOW to study effectively and accountably.
How we built it
LockedIn was built as a real-time, voice-first tutoring system. The browser frontend connects through LiveKit using a token from a small Node.js server and streams microphone audio in real time. A Python voice agent listens to the user, transcribes speech with Deepgram, generates responses using Google Gemini, and speaks back using Google Cloud Text-to-Speech. LockedIn includes focus tools like Pomodoro sessions, where the user's screen will be paused during break, and will ask them to recap the session to help stay engaged while studying.
Challenges we ran into
At the beginning, getting the LLM to work well with voice was harder than expected. It would sometimes cut itself off mid-response or hallucinate when context was incomplete, and those issues are way more obvious when the model is speaking instead of typing. Voice forced me to be much stricter about prompts, response length, and how the model handled interruptions.
LiveKit itself worked well, but debugging real-time audio was tough because problems often showed up in the wrong place. A bad token or SDK mismatch could look like a microphone or connection issue in the browser, even when the real bug was on the server. Getting everything stable meant carefully tracing audio and auth across the frontend, backend, agent, and LiveKit Cloud.
Accomplishments that we're proud of
Building LockedIn solo in a 24-hour hackathon was a real challenge. I tried to push as many core features as possible into a short time window, knowing I wouldn’t be able to build everything I envisioned. While the product isn’t fully polished, it works, and it clearly demonstrates the main ideas behind LockedIn. Given the time constraints, I’m proud of what I was able to ship and demo.
What we learned
One of the biggest takeaways was learning to trust myself and my ideas. There will always be doubts, moments where you question whether you’re on the right path, but sometimes you just have to go all in and believe. Hearing from inspiring people at NexHacks reinforced that mindset. Being in college and in my 20s feels like the best time to try things, take risks, and see what happens, whether it ends in success or failure. If I’m going to give something a shot, I might as well fully commit, knowing I can always adjust or step back later.
What's next for LockedIn
This is just the beginning for LockedIn. Next, I want to add post-session summaries so students walk away with clear, structured notes from what they just learned. I also plan to introduce an optional camera-based presence check through Overshoot to encourage accountability and reduce disengagement.
On the engagement side, I want to experiment with different voice modes, such as Gen Z–style voices or varied accents, to make studying more interactive and cognitively engaging. Research suggests that slightly increasing listening difficulty can encourage deeper processing and improve retention, and I’m excited to explore how this can be applied to learning.
Looking ahead, my long-term goal is to partner with standardized test prep platforms to distribute LockedIn at scale and bring real-time, psychology-backed studying to more students.
Log in or sign up for Devpost to join the conversation.