InterviewLens
AI that researches how your target company actually interviews โ then simulates it. No generic questions. Google is not Amazon. Amazon is not Toss.
Case Study
The Problem
Generic interview prep tools ask the same questions to everyone. But Google interviews differently than Amazon, which interviews differently than Korean startups like Toss. Candidates were preparing for the wrong interview.
My Approach
Split the hiring process into five agents on Gemini 2.5 Flash Lite, each with one narrow, focused job. Narrow agents are easier to prompt, test and debug than one giant prompt, and a failure in one stage never corrupts the others. Every AI call is wrapped in exponential backoff retries with layered fallbacks, and candidate memory lives in Firestore so it compounds across sessions.
The Result
A live, company-specific interview simulator that remembers returning candidates and survives failed AI calls without crashing. Resume-aware via Gemini multimodal PDF parsing. Built solo, end to end in TypeScript, and submitted to Google Gen AI Academy APAC โ Meet the Builders.
5-Agent Pipeline
- 01
Research
Studies how the target company interviews โ rounds, what it rewards, what gets candidates rejected
- 02
Interviewer
Plays a real interviewer from that company and asks culturally specific follow-ups
- 03
Evaluator
Scores each answer against the company's own criteria
- 04
Round Summary
Condenses each round into strengths and gaps
- 05
Final Verdict
Delivers the hiring-style verdict with improvements and cultural tips
Engineered for Reliability
Narrow agents
One job per agent keeps prompts small, testable and easy to debug.
Retries & fallbacks
Exponential backoff plus layered fallbacks โ a failed AI call never crashes the session.
Compounding memory
Candidate memory in Firestore makes every session more personalized than the last.
Multimodal input
Gemini parses the uploaded resume PDF so questions match the candidate.
What It Does
- 5-agent pipeline simulates a full company-specific hiring process end to end
- Research agent studies how the target company actually interviews
- Interviewer agent asks culturally specific follow-up questions
- Evaluator, Round Summary and Final Verdict agents score, summarize and decide
- Resume-aware via Gemini multimodal PDF parsing
- Candidate memory compounds across sessions via Firestore
Challenges & Solutions
Challenge 1
Most prep tools give generic questions โ but every company interviews differently
Solution
5-agent pipeline: Research โ Interviewer โ Evaluator โ Round Summary โ Final Verdict, each agent with one narrow, focused job
Challenge 2
A single failed AI call could crash a live interview session halfway through
Solution
Exponential backoff retries with layered fallbacks, so a failed AI call degrades gracefully instead of crashing the session
Challenge 3
Returning users started from zero every time
Solution
Candidate memory stored in Firestore compounds across sessions, so evaluation gets more personalized over time
Project Gallery
What I Learned
Multi-agent orchestration with narrow agent roles, retry and fallback design for LLM calls, persistent agent memory, multimodal PDF parsing with Gemini