NCHRecruitPro needed AI capabilities across 15+ domains — from resume parsing to org chart recommendations. Instead of building a monolithic AI service, we designed a multi-agent architecture where each agent specializes in one task.
Why Specialized Agents
A single large prompt handling everything fails in predictable ways: context window overflow, inconsistent output formats, and impossible debugging. Specialized agents with their own tools, prompt templates, and output parsers solve all three problems.
The Agent Coordinator Pattern
Complex workflows like candidate evaluation require multiple agents working together. Our Agent Coordinator orchestrates these workflows — it decides which agents to invoke, passes context between them, and aggregates results. For candidate screening, it runs resume matching, culture fit analysis, and skill gap assessment in parallel, then combines scores.
Cost Optimization Strategy
Running 40+ agents on GPT-4 would cost thousands per month. Our fallback hierarchy uses Groq (llama-3.3-70b) as the primary provider — its free tier handles 90% of our inference. Gemini serves as secondary, OpenAI as tertiary. Total AI cost: under $12/month for 2,400+ assessments.
Monitoring with LangSmith
Every agent invocation is traced through LangSmith — we can see exact prompts, responses, latencies, and token usage. When an agent produces unexpected output, we trace back to the exact prompt that caused it.
Results
The platform processes 500+ resumes daily with sub-second response times. Culture fit scoring provides data-driven candidate assessments that reduced subjective hiring bias. The entire system runs on a single VPS with Docker and Nginx.