01In short
Recruiters spend hours reading résumés and running first-round interviews before they meet anyone seriously. HireAI automated both steps.
In plain terms: I taught an open-source AI model (Llama 3.1) to rank résumés the way experienced recruiters do, using a lightweight fine-tuning method called LoRA. Then I built an AI interviewer that asks candidates questions and follows up on their answers in real time.
02What I did
- Fine-tuned Llama 3.1 with LoRA for résumé ranking — its rankings agreed with human recruiters 28% more closely than GPT-3.5 and Claude.
- Built a real-time AI interviewer that writes interview questions on the fly and asks context-aware follow-ups based on what each candidate says.
- Cut recruiter screening time by 80%.
- Signed 32 clients and generated $12K in revenue.
03How the ranking was measured
The goal wasn't for the AI to invent its own idea of a good candidate — it was to agree with experienced recruiters. After fine-tuning, its rankings matched theirs 28% more closely than GPT-3.5 and Claude did.
04Why I started it
Early-round hiring is slow and repetitive. Recruiters read hundreds of résumés and run nearly identical screening calls before they ever meet the people they actually want to hire.
General-purpose AI models could already read a résumé, but they didn't rank candidates the way experienced recruiters do. HireAI was built to close that gap, between a model that sounds reasonable and one that agrees with the people making the decision.
05Building the interviewer
The interviewer doesn't read from a fixed script. It generates questions for each role and candidate, listens to the answer, and decides what to ask next: digging into a vague answer, or following up on something interesting, the way a good human interviewer would.
Together with the résumé ranking, it took over most of the first round of screening and cut recruiter screening time by 80%.
The hard part wasn't generating questions. Language models are good at that. It was judgment: knowing when an answer deserved a follow-up and when to move on, so the interview stayed within time and still covered everything the role needed. Getting that right is what made it feel like an interview rather than a questionnaire.
It also had to be fast. In a live conversation, a pause of a few seconds feels broken, so the questions had to come back quickly enough to keep the conversation flowing.
06Running a company
Running HireAI meant much more than building models. I had to find customers, price the product, and support 32 clients, which brought in $12K in revenue.
The biggest lesson was that the technology only matters if it fits how people already work. Success wasn't a benchmark score. It was whether recruiters trusted the rankings enough to change how they spent their time.