Early-stage B2B counterparty screening requires analysts to manually cross-reference corporate registries, regulatory lists,and lengthy legal terms. This fragmented process takes hours per vendor and is prone to oversight. At the same time,deploying a standard LLM for compliance creates a critical risk: models default to a "helpfulness bias" and fill missing registry data with plausible-sounding hallucinations.
I designed and built a custom AI intelligence agent that standardizes initial vendor screening into an instantaneous,structured Zero-Trust snapshot, isolating ambiguous findings for human review.
Role: Context & AI Workflow Architect
Stack: OpenAI Custom GPT, System Prompt Engineering, Guardrail Architecture, Comparative Evaluation
Zero-Trust Boundary Conditions: Explicitly prohibited the model from inferring unverified corporate facts or assuming compliance based on brand prominence. When primary registry data is missing, the agent must label the field FLAGGED/UNKNOWN instead of guessing.
Mandatory Human-in-the-Loop (HITL) Routing: Enforced strict negative constraints barring the agent from issuing legal advice, regulatory clearances, or vendor approvals. Unverified items automatically populate an actionable checklist for a compliance officer.
Adversarial Edge-Case Testing: Benchmarked the agent using comparative entities, contrasting transparent companies (such as Stripe) against opaque, multi-jurisdictional structures (such as Telegram across Dubai and the BVI). When initial tests revealed the model hallucinating filings based on Telegram's consumer visibility, I re-engineered the instruction hierarchy to eliminate assumptions.
Rollout & Governance Blueprint: Designed an adoption framework for a 30 to 50 person operations team, anchored in live edge-case workshops and tracked against three KPIs: ≥40% faster pre-audit turnaround, >80% team adoption, and zero false negatives on high-risk escalations.
Live Prototype: B2B Pre-Audit & Compliance Intelligence Agent (Custom GPT)