Skip to content
Back to selected work

Case study

Recruitimate

An AI-native hiring platform that screens resumes, runs structured interviews, and scores candidates on what they actually say.

My role
Core engineer at Arken Systems, owning features end to end.

The problem

The problem

Hiring teams drown in resumes, and the AI interview tools that promise to help by scoring faces and tone are scientifically weak and now banned in EU hiring.

The constraint we chose

The constraint we chose
Candidate scoring is grounded strictly in interview transcripts, with no facial or vocal affect inference, aligned with the EU AI Act.

Scoring is grounded strictly in transcript content and excludes facial and vocal affect inference, keeping outputs consistent, explainable, and aligned with emerging AI hiring regulation such as the EU AI Act.

That refusal is a product decision, not a disclaimer. It is what makes the scores explainable to the hiring teams who act on them. I wrote about the reasoning in We Taught Our Hiring AI to Ignore Your Face.

What I built

What I built

The platform

A Next.js and TypeScript frontend and a FastAPI backend on a multi-tenant PostgreSQL architecture, with features owned end to end.

Real-time interviews

Multi-step interview intelligence workflows combine real-time audio over LiveKit (WebRTC), streaming transcription with AssemblyAI, and structured LLM scoring through tool use and function calling.

Prompt pipelines

Token-efficient prompt pipelines across the Claude and Gemini APIs cover system prompt design, conversation history and context window management, streaming responses, and cost-aware model selection.

Resume screening

Retrieval-augmented resume screening runs an embedding-based retrieve-then-rerank pipeline on PostgreSQL with pgvector, cutting manual shortlisting effort by roughly half in internal trials.

Outcome

Outcome

~50%

less manual shortlisting effort in internal trials

Stack

Stack

Next.jsTypeScriptFastAPIPostgreSQLpgvectorLiveKitAssemblyAIClaude APIGemini API