Risk intel
Ethical supplier audit
Autonomous procurement agent that turns web signals into a 0–100 ESG risk scorecard with deterministic scoring.
Marwa Bouabid & Kat Zhang
Why this project
We wanted to find out about Agentic AI. It sounded like a buzzword, now we know what agents do. The answer is everything and anything automatically and independently.
Demo
Project summary
Autonomous risk-intel agent for procurement that crawls fresh web results, extracts structured incidents via GPT-4o, and computes a normalized 0–100 ESG risk score using deterministic Python logic.
Built around a Streamlit UI with a gauge and expandable incidents. SearchAgent issues focused Tavily queries (labor violations, environmental fines, supply chain scandals), dedupes URLs, and passes snippets to a structured-output LLM parser with Pydantic validation.
Highlights
- Strict schema (SupplierAnalysis → RiskEvent) enforces severity (1–5) aligned to business rules.
- Recency decay buckets (≤6mo 1.0x, 6–12mo 0.8x, 1–2y 0.5x, >2y 0.2x; unknown 0.5x) to weight incidents.
- LLM never does math: Python handles scoring and caps at 100 via (sum/15)*100.
- Streamlit UI with Plotly gauge and incident list plus graceful handling of empty/failed searches.
Business problem
Give procurement teams an automated ESG early-warning system so they can spot supplier risk (labor, environmental, governance) without manual research or subjective scoring.
Results
Findings: Deterministic scoring produced consistent 0–100 risk ratings; recency weighting surfaced recent labor disputes over older minor infractions; schema enforcement prevented hallucinated fields.
Business outcome: Faster supplier diligence and clearer escalation thresholds, enabling proactive vendor management and fewer surprises in audits.
Tech stack
- Python 3.10+
- Streamlit
- LangChain / GPT-4o
- Tavily
- Pydantic
- Plotly
Analytics & methods
- Structured LLM extraction
- Recency decay
- Deterministic scoring
- Severity ladder
- Gauge visualization
Download a PDF with all references used for this project.
Download references