Multi-agent AI pipeline for engineering talent acquisition insights.
Author: Lamonte Smith | Senior Software Design Release Engineer, General Motors
Program: Interview Kickstart Applied Agentic AI - Capstone 2 | April 2026
GitHub: github.com/LSmithPMP
Engineering managers and talent teams lack real-time visibility into hiring pipelines. This system deploys 9 specialized AI agents delivering actionable insights in under 50 seconds for less than one cent per run.
| Layer | Technology |
|---|---|
| Agent Framework | LangChain |
| LLM Provider | OpenAI (gpt-4o-mini, gpt-4o) |
| Vector Store | ChromaDB |
| Orchestration | n8n Cloud |
| API Layer | FastAPI + Uvicorn |
| Dashboard | Streamlit |
| Data Contracts | Pydantic v2 |
| Agent | Autonomous | Role |
|---|---|---|
| RoutingAgent | Yes | Selects gpt-4o-mini vs gpt-4o per task complexity |
| EvaluationAgent | No | LLM-as-judge quality gate |
| OptimizationAgent | Yes | Autonomous cost and threshold decisions |
| Agent | External Tools | Role |
|---|---|---|
| SourcingQualityAgent | None | Channel conversion rates, cost per hire |
| RejectionPatternAgent | None | Stage bottlenecks, JD mismatch patterns |
| PanelLoadBalancerAgent | None | Interviewer overload detection |
| OfferInsightsAgent | None | Offer decline analysis, compensation gaps |
| PipelineHealthAgent | None | SLA breach analysis, velocity metrics |
| MarketIntelligenceAgent | Web Search API | Real-time market comp benchmarks |
Every agent returns: recommendation - evidence - confidence_score - cost_of_insight - alternative
- Trigger: POST webhook after every pipeline run
- Nodes: Receive Pipeline Results → All Agents Passed? → Calculate Pipeline Health Score → Format Success/Failure Response
- Trigger: Scheduled every 6 hours
- Nodes: Every 6 Hours → Load Pipeline Data → Analyze Critical Alerts → Requires Action? → Format Report → Consolidate
- Alert types: SLA_BREACH (CRITICAL), OFFER_DECLINE (HIGH), MARKET_GAP (HIGH), LOW_CONFIDENCE (LOW)
| Metric | Value |
|---|---|
| Agents passing | 7/7 (100%) |
| Eval score range | 0.60 to 0.97 |
| Cost per run (mixed) | $0.007404 |
| Latency | 25 to 46 seconds |
| Golden dataset | 75% pass rate (15/20) |
| n8n notification | 200 OK |
| Lever | Before | After | Savings |
|---|---|---|---|
| Model routing | All gpt-4o ~$0.040/run | Mixed $0.007/run | ~83% |
| Few-shot prompts | PipelineHealth 0.60 | PipelineHealth 0.97 | +62% quality |
pip install langchain langchain-openai langchain-community langchain-chroma
pip install chromadb streamlit pandas pydantic openai fastapi uvicorn python-dotenv requestsCreate .env:
OPENAI_API_KEY=your_key_here
LANGSMITH_API_KEY=your_key_here
LANGSMITH_TRACING=true
python3 orchestrator.py
streamlit run dashboard/app.py
python3 api.pySenior Software Design Release Engineer - General Motors Advanced Infotainment, Compute & Connectivity - Milford, MI