Every placement season, the same thing happens on campus.
A student walks into the TCS or Wipro HR round knowing every DSA answer cold, and then freezes on: "We have a 2-year bond and posting could be anywhere in India — are you okay with that?"
That's not a technical question. It's a trap. And nobody trains for it, because there's no cheap, honest way to practice it — the only feedback most students get is a rejection email, or a senior's half-remembered war story from two years ago.
I've spent the last few months building the thing I wish existed when I was that student: an AI that actually interviews you, live, in your voice, and tells you the truth about how you did.
Why I even started this
Before this, I built ApexGrab. That project taught me something I didn't expect it to: most "AI-powered" tools people ship are really just a chatbot wearing a costume. Type a question, get a text answer, call it intelligence. It's fine for a lot of things. It is not fine for interview prep, because the entire skill you're trying to build — staying composed, thinking on your feet, holding eye contact while your brain is short-circuiting — doesn't show up in a text box. It shows up under pressure, in real time, out loud.
So the bar for this one was different from day one: it had to talk, it had to listen, and it had to watch.
The stack nobody warns you about
The idea is simple to say and genuinely hard to ship: a live video/audio AI interviewer that responds like a person, not a script.
Here's roughly what's under the hood:
Google's Gemini Live API powers the actual conversational avatar — the part that listens, thinks, and responds in near real-time.
Pipecat handles the AI orchestration layer, stitching together speech-to-text, the model, and text-to-speech into something that feels like a conversation instead of a call-and-response loop.
Daily.co carries the WebRTC video/audio pipe, because building your own real-time media infrastructure from scratch as a solo founder is a fast way to burn a year.
MediaPipe watches posture and eye contact in the background, frame by frame, and turns "you looked away a lot" into an actual number instead of a vibe.
FastAPI on Railway and Next.js on Vercel run the actual product, with Supabase holding users, sessions, and progress data.
The genuinely hard part wasn't any single piece. It was latency. A human interviewer can pause for half a second and it reads as thoughtful. An AI that pauses for half a second reads as broken. Getting the whole pipeline — speech in, model thinking, speech out — to feel like a real back-and-forth took far more tuning than the actual "AI" part of the AI interviewer.
Building in the questions that actually matter
The technology was only ever half the problem. The other half was knowing what to ask.
Generic interview questions are everywhere. What's missing is the India-specific stuff that actually decides outcomes at campus placements: the bond clause question, the relocation question, the "what's your expected CTC" question that has no good textbook answer. I built the interviewer around these specifically — TCS, Infosys, Wipro, Accenture, Cognizant, HCL, Tech Mahindra, Zoho — because a generic "tell me about yourself" bot doesn't prepare anyone for the moment that actually trips students up.
It also remembers. Session to session, it tracks where a student is weak and turns the dial up there instead of repeating the same easy round forever. That part mattered more to me than the flashy avatar — repetition without adaptation is just expensive busywork.
Where it actually stands right now
I'm not going to pretend this is some polished unicorn story. It's live, it's real, and it's in the unglamorous phase every solo founder knows: getting the first real users, watching where they drop off, fixing what breaks, doing it again the next day. Founders' Circle pricing exists because early users take a risk on something unproven, and that risk should be worth their while.
What I can say honestly: watching a student go from stiff, one-word answers in session one to actually holding a real conversation by session three is the reason I keep building this instead of something easier.
What I'd tell someone starting the same thing
Build the boring infrastructure first. The avatar is the part people notice; the WebRTC pipeline, the auth, the payment retries are the part that decides whether any of it works at 11 PM when a student is trying to prepare the night before their interview. Nobody claps for that work. It's the whole product anyway.
If you're a B.Tech student prepping for placements, or just curious how a live AI interviewer actually holds up under real conversation, I'd genuinely like to know what you think: interviewphodo.com
Top comments (1)
Share Your feedback and let me know what you liked and what needs work?