Client projectAIProduction
MarqHire AI
Hiring platform that parses resumes, scores them against a job description, and runs first-round screening interviews by voice.
Result
Screening automated end to end
The problem
Recruiters hiring at volume spend most of their time reading resumes that were never going to match.
What I built
Built resume parsing that scores each candidate semantically against the job description rather than by keyword match, then added a voice agent to run the first screening call.
The outcome
A configurable pipeline teams can configure themselves, with candidates ranked live as applications arrive.
How it works
- 01
An application arrives and the resume is parsed into structured fields rather than stored as a file nobody opens.
- 02
The candidate is scored against that specific job description semantically, so a relevant background phrased in unfamiliar words still ranks.
- 03
Anyone above the bar gets a voice screening call from an agent working through the role's own question set.
- 04
The call is transcribed and summarised against the same criteria for every candidate, and the shortlist updates as applications land.
- 05
The recruiter moves candidates through a pipeline they configured themselves, with configurable stages, instead of one the tool imposed.
The AI layer
Embedding-based matching between resume and job description, so relevance survives paraphrase, plus a voice agent for the screening call and structured summarisation of what was said.
The engineering layer
Resume parsing across inconsistent file formats, a configurable stage machine, ranking that updates as new applications arrive, and telephony plumbing with the retry and fallback behaviour a real call needs.
Key technical decisions
Semantic scoring, not keyword filters.
Keyword filters reject the candidate who wrote the same experience in different words. Scoring against the job description in embedding space is the reason to do this with a model at all.
The screening call is the automation worth having.
Ranking a list still leaves someone making twenty phone calls. Automating the first call is where the hours actually go.