🚀 Struggling to Keep Up with Resumes and Calls? Let's Talk AI Engineering in Hiring
Hiring is broken. Well, not broken—just painfully inefficient.
CVs pile up. Your phone overheats from too many calls. And your 9-to-5? That’s turned into a 24/7 slog of reading resumes, conducting the same first-round questions, and chasing candidates.
That’s why SkillSociety exists—to automate the first stages of hiring using AI.
And because this is dev.to, let’s skip the fluff and get straight into the tech behind it.
💡 How We Built an AI Interviewer
SkillSociety is a fully automated AI-powered interview system. Instead of manually screening dozens or hundreds of candidates, our AI handles:
✅ Pre-qualification – Filter applicants based on topic-related question based on the real job. Here a Skill Society we do believe that CV are a thing of the past!
✅ AI Interviews – Candidates speak with an AI-powered interviewer in real-time.
✅ Intelligent Evaluation – AI scores responses based on role-specific and personalised criteria.
✅ Insights & Transcripts – Structured interview logs for hiring managers to review.
Now, let’s break down the tech stack that powers all this.
🏗 The Tech Stack
We designed SkillSociety to be scalable, real-time, and fully automated using a modern web stack:
- 🏗 Frontend: NextJS (with Server Components)
- ☁ Backend & Auth: Supabase (PostgreSQL + RLS)
- 🧠 AI: A combination of: OpenAI, LLama, Deepseek, Sonnet, Whisper, Deepgram, and who know what will be next (LLMs + TTS + ASR)
- 🚀 Edge Deployments: Vercel (Serverless Edge Functions)
🧠 AI Architecture: LLM + TTS + ASR + (Tooling to cover in another blog post)
AI interviews aren’t just a simple chat with ChatGPT. We needed to build a system that:
- Understands spoken responses.
- Evaluates candidates dynamically.
- Speaks naturally, with contextual follow-ups.
Here’s how we do it:
📜 Step 1: Large Language Models (LLMs) for Interview Logic
At the heart of our system there are multiple fine-tuned LLM, responsible for:
- Parsing job descriptions to auto-generate tailored interview questions.
- Evaluating candidate responses using structured scoring models.
- Asking follow-ups dynamically based on candidate answers.
LLM Fine-Tuning:
- We fine-tune the model on real-world hiring data to improve domain expertise.
- Prompt engineering ensures questions flow naturally and are job-relevant.
- We generate role-specific prompts, adjusting based on industries (tech, finance, healthcare, etc.).
🔊 Step 2: Text-to-Speech (TTS) for Human-Like AI Interviewers
A text-based chatbot isn't enough. We needed AI to speak, making the interview natural.
Why We Use TTS:
- Multi-language support for global hiring (We're in Australia and Italy at the moment!).
- Human-like responses for a smooth candidate experience.
- Different tones and accents for regional customisation.
🛠 Tech Used:
- Microsoft Azure – It's one of the cloud provider we find most attractive from the deployment point of view of new models.
🎙️ Step 3: Automatic Speech Recognition (ASR) for Candidate Responses
Instead of making candidates type answers, we process spoken responses with ASR.
ASR Features:
- Real-time speech-to-text transcription (low latency).
- Noise filtering (removes background noise).
- Handles different accents & speech patterns.
- Keyword extraction for automated scoring.
🛠 Tech Used:
- Whisper ASR (OpenAI) – Transcribes speech to text.
- Deepgram – Backup model for multi-accent detection.
⚡ Real-Time Interview Pipeline
Here’s how the AI Interview works under the hood:
-
Candidate starts an AI interview.
- 🟢 LLM generates role-specific interview questions.
- 🟢 TTS converts questions into speech.
-
Candidate answers in real-time.
- 🔵 ASR transcribes spoken response.
- 🔵 LLM analyses the answer & generates follow-ups.
-
AI dynamically adjusts interview flow.
- 🟡 Scores the candidate based on predefined hiring criteria.
- 🟡 Auto-generates summary reports for recruiters.
🛠️ Engineering Challenges
Building an AI-driven interview platform wasn’t easy. Here are a few challenges we tackled:
⚡ Low-Latency AI Processing
Nobody wants long pauses between AI questions and candidate responses.
We optimised performance by:
- Caching LLM calls for common follow-ups.
- Running ASR inference on the Edge to cut down response times.
- Parallelising API requests between LLM, TTS, and ASR.
🏎️ Result: Sub-2s response times for dynamic interviews.
DrawBack: We had to create a turn-taking mechanism that is different from anything else in the market. Our use case is Q&A, however with a grain of salt, as the candidate can take some seconds to think about the answer, and you don't want the AI always interrupting or speaking on top of them.
🏗️ Scalability & High Concurrency
Hundreds of interviews can happen simultaneously.
To prevent bottlenecks, we:
- Deployed ASR on a dedicated GPU cluster.
- Rate-limited free-tier users to prevent abuse.
🏗️ Result: No slowdowns, even with high traffic.
🔒 Data Privacy & Compliance
Recruitment data is sensitive. We built SkillSociety with:
- Full encryption (PostgreSQL + Row-Level Security).
- GDPR & ISO 27001 compliance for enterprise users.
🔐 Result: Secure, enterprise-ready AI hiring.
🔥 AI-Powered Hiring: Built for Devs, by Devs
If you’re a hiring manager, recruiter, or startup founder tired of manual screening, it's time to let AI do the grunt work.
SkillSociety automates first-round interviews with:
✔ Real-time, voice-based AI assessments
✔ Scoring models tailored to each job
✔ Fully customisable hiring workflows
🔗 Check out SkillSociety and see how AI-powered hiring actually works.
🚀 AI isn’t replacing hiring managers—it’s eliminating the boring parts.
Top comments (0)