senuka.dev

#7skillhunt.md

SkillHunt

An AI interview platform where candidates answer out loud and an LLM grades the transcripts.

Origin
Tecciance
Role
Full-Stack Developer
Period
4 months

overview.md

SkillHunt is an AI interview platform at skillhunt.net, built at Tecciance. A headhunter enters a job description, an LLM generates the interview questions, and SkillHunt gives back a unique interview link.

The candidate answers the questions out loud. The browser records the audio and takes periodic screenshots as an anti-cheating check. A mock interview mode lets candidates practise first.

Surface-level view on what I worked on

what-i-built.md

Audio and screenshots upload in chunks through AWS Lambda, with S3 and API Gateway behind it, and each chunk is retried if it fails.

A Lambda also transcribes the audio.

An LLM grades the transcripts. It is OpenAI today and the model can be swapped, and I researched cheaper models for smaller operations. I tuned the prompts and helped write the grading algorithm.

Bulk grading runs on a Spring async pool.

Grading results are stored in PostgreSQL, which feeds the report and email views.

I also built the Next.js report and email views.

what-it-changed.md

  • +40%Bulk throughput
  • +15%Model accuracy
  • −20%LLM latency
  • −25%Serverless cost

Prompt tuning raised grading accuracy by 15% and cut latency by 20%.

The async pool raised bulk grading throughput by 40% and took processing from minutes to seconds.

The serverless module cut infrastructure cost by 25%.

Stack

stack.diff

  • TypeScript
  • Next.js
  • Java
  • Spring Boot
  • PostgreSQL
  • OpenAI
  • AWS EC2
  • Lambda
  • S3
  • API Gateway
  • CI/CD
  • LLM prompting