Hi all,
I built AEGIS as my Moonshot Project for DSU’s NVIDIA RTD (Ready to Deploy) 2026 Boot Camp — a banking complaint-resolution system where NVIDIA NIM handles response drafting, but nothing NIM writes reaches a customer unverified.
The pipeline: a Spring Boot orchestrator classifies each complaint (fine-tuned DistilBERT, 11 CFPB categories), runs a compliance check, retrieves relevant regulation text via pgvector hybrid search (94% hit-rate), then NIM drafts a response grounded in that retrieved context. Before a human ever reviews it, a grounding gate checks the draft against the original complaint and the retrieved text — invented amounts, fabricated phone numbers, and unsupported claims get blocked, not shipped. It caught a real hallucination during development; there’s a regression test for that case now.
PII is redacted before any prompt reaches NIM. Every response is signed off by a human in a maker-checker console before sending, with an immutable audit trail.
Evaluated on a temporal holdout pulled fresh from the live CFPB complaints API (not a curated demo set): 77.3% classification accuracy across the 11 categories, 0.93 F1 on escalation detection. Load-tested to ~63 req/s sustained.
Happy to go into more detail on the grounding-gate design or the NIM integration if anyone’s curious.