Chay Patil C2X ยท Context-to-Execution Let’s talk

Work/Peoplebox.ai/YC S22

Teaching an AI recruiter to read roles

Lead, Data Management & Annotation, Feb 2025 to Jan 2026. I led a five-person team on the data, review and fine-tuning loop behind Peoplebox's AI recruiter, from planning to production in under 3 months.

[ Back ]

01/The problem

An AI can't match a candidate to a role it can't read.

“SDE-3, Platform” and “Sr. Backend Engineer (Go)” are close to the same job. “Staff SRE” and “Junior QA Automation” are nowhere near each other. Before the recruiter could match anyone, the model had to understand domain and seniority the way a good recruiter does. Engineering roles are the example here, the work covered more.

From raw job titles to a model that matches candidates

Roles in · tagged · reviewed

    The model

    Collecting labelled roles

    Model is guessing

    Candidate: backend engineer, 6 years, Go

    Role: Sr. Backend Engineer (Go)

    Strong match

    Example roles. The 98% is the real result after recurring fine-tuning; the climb shown here is illustrative.

    02/The loop

    Five steps, run again and again until it was right.

    The model only got better because the data behind it got better. Every round tightened the labels, ruled on the edge cases and fed it back in.

    1. 01Data

      Pull in the roles

      Large sets of real job roles, the messy titles companies actually write.

    2. 02Human

      Tag every role

      Annotators label each one by domain and seniority against a shared rubric.

    3. 03Review

      Review and rule on edge cases

      Every label checked. Tricky cases decided once, written down as rules, reused.

    4. 04AI

      Fine-tune and test

      Reviewed data trains the model. Accuracy measured every round.

    5. 05AI

      Match candidates to roles

      The recruiter predicts who fits the roles clients need filled.

    6. 06Feedback

      Clients close the loop

      Beta feedback from clients came straight back into the labelling rules.

    AI does the workPeople decidePlumbing

    03/The result

    Shipped to real hiring teams in under 3 months.

    Clients like Hasura and Razorpay used it to hire. I reported daily to the CEO and CTO, working alongside product and engineering.

    98%Accuracy after fine-tuning, 5% above Gemini 2.5 Pro at the time
    <3 moPlanning to production
    5People on my team

    Next/Your data

    Need an AI that actually knows your domain?

    [ Let's talk ]