The experiment tested agentic search capabilities using open-source models to identify HR and AI/ML specialists at Sber. The researcher utilized a stack including Ollama for local models, OpenRouter for model access, and a custom 'JEV' judge model to filter results. The total budget allocated was $20, split between Ollama and OpenRouter services.
Initial tests revealed that mainstream search APIs like Google and Bing are either restricted, costly, or prone to CAPTCHA blocks. Consequently, the search process relied on Yandex via SearXNG. The researcher found that adding complex 'skills' to models increased processing time from 52 to 121 seconds while sometimes reducing recall, suggesting that a simple model paired with a fast filter is more efficient.
Testing six open-source models showed that while finding HR contacts proved difficult, models like gemma4:31b, nemotron-3-super, and gpt-oss:120b successfully identified AI/ML specialists. Performance varied significantly between runs, with gpt-oss:120b demonstrating the most stability and speed. Nemotron-3-ultra was the slowest and most expensive, costing up to $0.38 per run.
Paid frontier models were also evaluated. DeepSeek and MiMo showed higher success rates in locating HR contacts compared to open-source alternatives. MiMo emerged as the most cost-effective solution in its round, costing $0.10 for four runs. Conversely, Gemini 3.8 flash proved to be the most expensive, with a cost of $0.0625 per identified contact.
