How Indeed uses AI to help evolve the job search

Core Finding

Indeed avoids monolithic LLM reliance, instead deploying a hybrid AI architecture: GPT-4 (via OpenAI API) handles high-semantic-complexity tasks (e.g., job description generation, interview question suggestion); Llama-3-70B and Llama-3-8B—fine-tuned on Indeed’s resume corpus—serve cost-efficient batch inference (e.g., resume summarization, skill extraction); and the internally built Ranking Model v3—a multi-task deep ranking model trained on real-time behavioral feedback—acts as the central matching engine.

Key Deployment Scenarios

• Job Matching: Ranking Model v3 integrates 12 signal types (including clickstream, dwell time, application funnel, cross-device behavior), delivering +27% CTR in Q2 2024 and +41% long-tail job exposure;
• Resume Parsing: A fine-tuned Llama-3-8B model achieves structured field extraction (skills, years of experience, education) on 92.3% of resumes with F1-score 0.89, delivering 3.2× higher throughput than legacy rule-based engines;
• Interview Assistant: A GPT-4-turbo–powered conversational tool provides personalized interview prep (including role-specific STAR question banks); average user preparation time decreased by 38% within 3 months post-launch;
• Employer Insights: Built on Hugging Face Transformers + custom temporal feature encoders, this module delivers real-time talent pool heatmaps and competitive job attractiveness benchmarks (aligned with Glassdoor and LinkedIn data sources), now deployed across Indeed’s Top 500 employer customers.

Technical Governance Practices

• All generative AI outputs are watermarked with ‘AI-assisted’ and include editable fallback interfaces;
• Resume parsing outputs undergo SOC 2 Type II–certified validation; PII fields (e.g., ID numbers, addresses) are automatically redacted;
• Ranking Model v3 is updated daily via incremental training (<15 min latency); its A/B testing framework implements causal inference rigor per arXiv:2305.16724 (‘Causal Inference for Online Experimentation’).