~/project/honkhonkhire.md
HonkHonkHire
AI mock interviews tailored to you
- Timeline
- September 15th - 17th, 2023
- Type
- Hackathon
- Team
- Donald Lee, with Vaibhav Sharma, Michelle Kim, Maureen Luo
- My role
- Full-stack. I connected the front end and back end and got Cohere working.
- Hack the North 2023
- ๐ Best Use of MongoDB Atlas
## overview
It's job search season (again)! One of the most nerve-wracking parts of applying for a job is the behavioural interview, and there isn't a great way to practice for it.
HonkHonkHire is an AI interview practice app. It reads your experience from your resume, LinkedIn, or portfolio, asks you questions about your own experiences, and gives you feedback based on your answers and facial expressions. We built it at Hack the North 2023 and won Best Use of MongoDB Atlas!
## why-we-built-it
Why we built it
Most practice tools can't really recreate an interview because their questions are generic. A real interviewer asks about the project on your resume, not a random question from a list. Interviewers also rarely give useful feedback, so you leave confused and without a plan to improve.
So we focused on two things: questions about your own experience, and feedback you can actually use.
## what-it-does
What it does
You upload your resume or link your LinkedIn or portfolio. OCR Space pulls the text out, and Cohere uses it to write interview questions about your experiences. While you answer on camera, MediaPipe tracks your facial expressions and your answer gets transcribed. Cohere then uses both to give you feedback on what went well and what to work on.
## what-i-worked-on
What I worked on
I was the full-stack developer on the team. I connected the front end and back end, passed data between them with Node.js, and got Cohere generating the questions and feedback. We stored resumes in Firebase and account settings in MongoDB.
It was my first time using Cohere and OCR Space, so I spent a lot of the weekend reading docs.
## getting-useful-questions-out-of-cohere
Getting useful questions out of Cohere
My first prompts gave back generic questions, the same kind you'd find on any list online. That was the exact problem we were trying to fix. So I gave Cohere more context. Along with your resume, the prompt includes what you told us during onboarding: your industry, the role and level you're applying for, and what you want to get better at. The questions started sounding like they came from someone who had actually read your resume.
The other problem was the format. Cohere's answers came back in a slightly different format every time, and the app would break trying to read them. So I asked for exactly seven questions in a fixed JSON format, which the app could read every time. Generating them still took a few seconds, so I added a loading page so the wait didn't look like the app had frozen.
## challenges
Challenges
We had a lot of ideas and not much time, so we limited brainstorming and looked at the sponsor prizes to help us choose. We also picked a backup idea in case we couldn't decide within a couple of hours.
The project was ambitious for a weekend, and we'd never worked together before, so we spent some time figuring out each other's strengths and how we liked to communicate. We only slept around three hours and planned our sleep so the right people were awake at the same time.
## key-learnings
What I learned
- 01
The quality of an AI feature depends on the context you give it. Generic input gets generic questions.
- 02
If code is going to read the model's answer, ask for a strict format. Treat it like an API, not a chat.
- 03
On a new team, figuring out each other's strengths early saved us more time than any shortcut.
## built-with
Languages
- Node.js
- JavaScript
- HTML
- CSS
Tools & libraries
- Cohere API
- OCR Space
- Google MediaPipe
- MongoDB Atlas
- Firebase
- Figma
## next-project
TravoAI
AI trip planner for you and your friends
read the write-up