AI Interview PlatformAlgorithmic question generation, automated candidate evaluation, 24/7 screening.

The challenge

A hiring team was spending most of its week on early-stage screening calls that produced little signal. They needed an interview process that ran around the clock, gave structured feedback, and surfaced top candidates without scheduling overhead.

What I built

Built an AI-driven interview platform with algorithmic question generation, automated candidate evaluation, and real-time data processing. Scalable REST APIs with RBAC, an integrated applicant tracking system, multi-language support, and scoring algorithms with weighted evaluation and rule-based decision logic.

• Algorithmic question generation tailored per role
• Weighted scoring + rule-based decision logic
• Applicant tracking system with multi-language support
• Parallel evaluation pipeline running 24/7

Key technical decisions

The choices behind the choices — and what tradeoff each one made.

• Rule-based scoring on top of LLM signals

  Pure LLM scoring is hard to audit and inconsistent across roles. Weighting LLM signals into a deterministic rubric per role gives recruiters something they can defend in a hiring decision.

• Parallel evaluation over a queue

  Candidates shouldn't wait for a sequential pipeline. A worker pool with per-candidate isolation means a slow API call on one interview doesn't block the rest of the cohort.

• ATS as a thin layer, not the product

  Recruiters already have ATS habits — competing with established tools is a losing fight. The platform integrates rather than replaces, which made adoption a soft sell.

Outcome

Screening time dropped 60%. Interviews run 24/7 without scheduling conflicts thanks to time-complexity-aware processes and parallel evaluation. Recruiters now spend their day on the shortlist, not the inbox.

Lessons & what I’d do differently

• Bias surfaces fast in scoring algorithms — bake calibration runs and audit logs in from version one, not as a future feature.
• Recruiters trust ranked shortlists more than raw scores. The UI shift from "7.4/10" to "top 5%" changed the perceived value.
• Multi-language support is half UX, half evaluation rubric — translating questions isn't enough if the scoring weights weren't tuned per language.